<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Feiza Data on Medium]]></title>
        <description><![CDATA[Stories by Feiza Data on Medium]]></description>
        <link>https://medium.com/@feizadata?source=rss-5da05b3c7452------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*vuUuxjyPmwunG7F_18nHWw.png</url>
            <title>Stories by Feiza Data on Medium</title>
            <link>https://medium.com/@feizadata?source=rss-5da05b3c7452------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 15:37:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@feizadata/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Word2Vec: The Geometry Behind Similar Words]]></title>
            <link>https://medium.com/@feizadata/word2vec-the-geometry-behind-similar-words-c1c87852ff0d?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/c1c87852ff0d</guid>
            <category><![CDATA[word2vec]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[word-embeddings]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 20:43:22 GMT</pubDate>
            <atom:updated>2026-11-01T20:43:22.932Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2YNy9pXJgqDGnA9DQVKaKw.png" /></figure><p>Type tortilla into a grocery search bar. A helpful search result might recommend similar products like pita, naan, or flatbread. This isn’t because somebody wrote down every possible similar product by hand. A model learned that those words are somehow similar in nature.</p><p>This idea is behind a lot of familiar experiences like related-search, product recommendations, or looking up synonyms. But before any model can suggest a related word, it must understand <strong>what “similar” actually means.</strong></p><p>For a computer, geometry is the answer. If we turn words into lists of numbers and map them in high-dimensional vector spaces, a model can decide which points (words) are closest. Even still, the method we use in such spaces to determine if two things are “close” can vary, and the definition we pick can meaningfully change results.</p><h3>Starting with the model behind the idea</h3><p>In 2013, Google created the <a href="https://arxiv.org/abs/1301.3781">Word2Vec</a> model to transform words into numbers that could relate to one another. More technically, Word2Vec is a two-layered neural network that learns a continuous numerical vector for every word in a text corpus it trains on. Each vectorized word, called a <strong>word embedding</strong>, captures how that word was used around other words in the corpus. A standard word embedding might have 100–300 dimensions that remain fixed after model training ends. That gives every word a position in a multi-dimensional vector space. As an example, vectorizing the words Banana and Strawberryproduce two unique word vectors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jdmKVJPIWGNF6aEwPMmvZQ.jpeg" /><figcaption>Each word is represented by a 100-dimensional vector in this experiment</figcaption></figure><p>Banana and Strawberry are close in similarity (they’re both fruits), and each respective vector is designed to pick up on this similarity. Training on a rich text corpus tunes those dimensions for each word.</p><p>But what does “closeness” really mean when words are represented as vectors with different directions and lengths? Instead of trying to picture a vector with hundreds dimensions, let’s pretend each word embedding only has 2 dimensions. We can plot each vector in two-dimensional space to see how close their coordinates appear.</p><p>Even in two dimensions, “close” has more than one valid definition. To consider a few, let’s look at Cosine Similarity, Euclidean Distance, and the Dot Product of any two vectors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GFfePPt-IdMo2bsxknfyeA.png" /></figure><p>After generating word embeddings from a trained Word2Vec model, most public libraries rely on cosine similarity to measure closeness between vectors (how similar two words are to one another).</p><p>I was curious to see how applying different distance metrics on the same set of word embeddings would impact the ability to flag word similarities.</p><h3>Setting up an experiment</h3><p>I found some retail product data from <a href="https://world.openfoodfacts.org/data">Open Food Facts</a> to create a grocery-style text corpus to train a Word2Vec model on. After filtering to US English product records only, deduplicating product titles, and removing brand names, each product name became one short training sentence. A single product, Apple &amp; Cinnamon Tea, was formatted as a list like [Apple, Cinnamon, Tea].</p><p>Next, I trained the Word2Vec model on this text corpus using the <strong>skip-gram</strong> method. It starts with one word in a sentence and predicts nearby words, which seemed reasonable for short product titles with specific food terms. After cleaning up the text corpus for training, it contained 538,138 product-title “sentences” (short lists of words from individual products) and 2.28 million tokens (words). It learned word embeddings for 14,460 words that appeared at least five times.</p><pre>model = Word2Vec(<br>    sentences,<br>    vector_size=100,<br>    window=5,<br>    min_count=5,<br>    sg=1,          # skip-gram method<br>    negative=10,<br>    sample=1e-3,<br>    epochs=20,<br>    seed=123,<br>    workeprs=1,<br>)</pre><p>I only trained <strong>one</strong> <strong>model</strong>, then ranked the same learned word embeddings in three ways (one per distance metric). Remember, Cosine Similarity only compares direction, Euclidean Distance compares the physical gap between vector tips, and Dot Product uses direction and vector length.</p><h3>How I scored similarity</h3><p>The scoring question: can a Word2Vec model identify close <strong>semantic peers</strong> of a word? Before checking any results, I picked 12 grocery terms and a small list of close peers I knew to be related to each term. For example, the term Tortilla was linked to similar peers I chose like Pita, Wraps, Naan, Flatbread, etc..</p><p><strong>Precision@5</strong> is a metric that measures how many of a model’s five closest suggestions are relevant. For each of the 12 grocery terms, I asked the model for its five nearest neighbors and counted how many of those recommendations appeared in the predefined peer list for that term. I then averaged the scores across all 12 grocery terms. If three of the five retrieved words were predetermined peers, that query scored 60%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4VxgHzVI-UMZYWapywVw5w.png" /></figure><h3>Results</h3><p>All three methods exceeded the random baseline, which reflected selecting five terms at random from the eligible vocabulary. <strong>Cosine similarity</strong> finished first, <strong>Dot Product</strong> was close behind, and <strong>Euclidean Distance</strong> was a clearer third place.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Pm294r-kuBObR5c2e73lw.png" /></figure><p>Cosine similarity found 43 close-peer hits across 60 retrieved slots. Dot product found 41 (a very small difference) and was a real contender for the text corpus used in this model.</p><p>Euclidean distance underperformed, likely due to its sensitivity around vector magnitude. I found that the <strong>frequency</strong> of a query word in the text corpus and the respective vector’s <strong>magnitude</strong> had a 90.8% correlation, showing that popular words tended to have longer vectors. Euclidean distance treats vector length differences as extra separation; dot product treats this length as a boost when vectors point in similar directions; and cosine similarity ignores vector length entirely. For this reason, frequency-related differences in vector magnitude may have hindered the effectiveness of Euclidean distance.</p><p>Cosine similarity performed best, suggesting direction was the clearer signal for product-name similarity. Dot product remained close (so magnitude wasn’t pure noise) but it still couldn’t emerge as the first place metric.</p><p>Let’s take a look at the top similar words retrieved using bread as the query word. This view stays at the word level: each column shows the five closest semantic peers ranked nearest to Bread by one metric. The query word itself is excluded, and candidates needed to appear at least 20 times to reduce spelling variants and rare-word noise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_6YUsrz2YmFcSktYmBjQPQ.png" /></figure><h3>Conclusion</h3><p>For this Word2Vec model, vector direction (cosine angle) was a cleaner signal than vector length (magnitude). The experiment helps explain why cosine similarity is a common default in public Word2Vec libraries. Cosine similarity compares the relationship between vectors without letting differences in length drive the ranking.</p><p>The question of how we define closeness extends beyond Word2Vec too. Search tools that leverage newer models like BERT or RoBERTa also turn language into vectors and rank results based on how close those vectors are. The experiment above does not establish which distance metric is best for <em>every</em> model or use case. Still, it shows that with any language embedding model, closeness is a choice worth testing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c1c87852ff0d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Symbolic Regression: A Math-First Approach to Extrapolation]]></title>
            <link>https://medium.com/@feizadata/symbolic-regression-a-math-first-approach-to-extrapolation-0df17f26c1d0?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/0df17f26c1d0</guid>
            <category><![CDATA[regression-modeling]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[symbolic-regression]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 17:48:25 GMT</pubDate>
            <atom:updated>2026-09-01T17:48:25.551Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*3wWTVJGg6RECMFTiR14fJg.png" /><figcaption>Thumbnail</figcaption></figure><p>While scrolling through LinkedIn a few weeks back, I stumbled across a <a href="https://www.linkedin.com/posts/nikos-gkkz_ml-datascience-ai-activity-7485340527131803650-HS80/">post</a> on a data model I’d never heard of before. Many traditional regression approaches require you to pick the shape of your model (linear, polynomial, exponential) before fitting it to your data. <strong>Symbolic Regression</strong> turns that around by searching for the model shape itself using the data it’s trained on.</p><h4>What makes Symbolic Regression different?</h4><p>To set the scene, consider a simple linear regression model with one input. The model itself takes a standard form like Y = β₀ + β₁X + ε, where ε is the error term. Training this model requires finding values for β₀ and β₁ that best fit the data. Visually, the model fits a straight line through a scatterplot of the predictor and response. Regardless of the beta coefficients chosen, <strong>the general shape of the model does not change. </strong>A linear regression model in this context will always produce a straight line, just with different slopes and intercepts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eHvNzCU2IObv89uaaDeylA.png" /><figcaption>A Linear Regression model plotted over training data, where X predicts Y</figcaption></figure><p>Symbolic Regression is different. It searches for a model shape by building equations that fit the training data. It tests combinations of different operators and functions (multiplication, exponents, trig functions, etc.) in search of equations that fit the observed training points more and more closely. With enough flexibility, that search can produce an equation that fits every point in the training data. Using the example above, a perfect fit can be derived with 5 terms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HzyJN0SGLNz_hpQwVzpZCA.png" /><figcaption>A perfectly fitted SR equation on the same training data</figcaption></figure><p>As someone who’s spent his fair share of time plotting crazy equations in <a href="https://www.desmos.com/calculator">Desmos</a>, having a model derive an equation like this sounded very cool. It’s the idea of building an equation that covers every point in a training data landscape, like collecting all the dots in a game of Snake.</p><h4>Isn’t this overfitting?</h4><p>It’s hard to ignore that this sounds like a textbook recipe for overfitting. We can derive an equation with near-perfect accuracy on known data, but that doesn’t guarantee strong predictive accuracy on new data.</p><p>This is where complexity comes into play. Every operator or function added to an equation increases the model’s <em>complexity</em>, and more complex equations can start to fit noise instead of signal. So, instead of optimizing for training accuracy alone, Symbolic Regression weighs accuracy against complexity. An elbow plot can be used to identify the point where adding more terms stops meaningfully improving the fit. Using this method, a simpler equation with 3 terms might be chosen instead of the perfectly fitted equation above.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fskykcDnp3kKtLnbidebCg.png" /></figure><p>From there, we arrive at a more practical Symbolic Regression equation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kqHYhcSyszE9f8mtK6mHrg.png" /><figcaption>An optimized SR equation fit to model training data</figcaption></figure><h4>Where is this actually useful?</h4><p>There are other constraints with Symbolic Regression worth mentioning. We’ve only discussed continuous variables so far. Say we wanted to use Symbolic Regression to determine the optimal pricing of retail goods. Pricing doesn’t only depend on continuous inputs like cost and demand. We’d probably want categorical features too, like store region or product type. Outside of getting fancy with some numerical encoding, Symbolic Regression isn’t the friendliest when it comes to handling categorical variables. It certainly isn’t a fit for <em>every</em> modeling problem.</p><p>Let’s look at one useful example for Symbolic Regression in the medical field. Many people with Type 1 diabetes use continuous glucose monitors (CGMs) alongside insulin pumps to manage their blood glucose throughout the day. CGMs don’t actually measure glucose directly from the bloodstream. They measure glucose in the interstitial fluid beneath the skin, which can lag behind blood glucose when levels are changing quickly. <strong>If we know how quickly a CGM reading is rising or falling, can we estimate the correction needed to the reading to better approximate true blood glucose?</strong></p><p>To test this, I generated a <a href="https://github.com/ethanfeiza/symbolic-regression-cgm/blob/main/cgm_data.csv">simulated dataset</a> of CGM readings, rates of change, and reference blood glucose measurements across multiple individuals. Using the <strong>CGM reading</strong> and its <strong>rate of change</strong> as inputs, I trained a Gradient Boosting model and a Symbolic Regression model to predict <strong>reference blood glucose</strong>. I held out some ordinary rates of change to test how well the models predicted within the range they had already seen, then used faster rises and falls outside the training range to test extrapolation. You can view the experiment code in <a href="https://github.com/ethanfeiza/symbolic-regression-cgm/blob/main/cgm_symbolic_regression.ipynb">this Python notebook</a>.</p><p>On held-out readings within the normal range, both models performed pretty well. Gradient Boosting had a Mean Absolute Error (MAE) of <strong>2.00</strong>, while Symbolic Regression came in at <strong>1.43</strong>. The bigger difference showed up when the models had to extrapolate to faster rates of change. Gradient Boosting’s MAE jumped to <strong>16.58</strong>, while Symbolic Regression held to <strong>7.60</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rZLH2juPXHN3Mwh5yAP1mg.png" /></figure><p>These results make sense when you consider how both models work. A Gradient Boosting model can struggle to extrapolate beyond the range of values it encountered during training. Symbolic Regression, on the other hand, leaves us with an explicit equation that can be evaluated beyond that range. In the experiment, we derived this equation for the Symbolic Regression model:</p><p><strong>Estimated Blood Glucose = (CGM Reading) + 7.4R − 3sin(R)</strong></p><p>where <strong>R</strong> represents the CGM’s rate of change.</p><p>This isn’t evidence that Symbolic Regression generally outperforms Gradient Boosting. The experiment is built around a continuous, structured relationship where extrapolation matters. This is merely a clean example where Symbolic Regression starts to make sense. The inputs are continuous, the relationship between them has some mathematical structure, and we care about what happens beyond the exact range of data we trained on.</p><p>Symbolic Regression still feels pretty niche compared with the models I’d reach for in most everyday problems. But that’s part of what makes it interesting. For the right problem, it can be surprisingly competitive. And the best part? It leaves us with a math equation!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0df17f26c1d0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Ultimate Tableau BAN Formatter Calc]]></title>
            <link>https://medium.com/@feizadata/the-ultimate-tableau-ban-formatter-calc-6908ac746836?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/6908ac746836</guid>
            <category><![CDATA[data-analysis]]></category>
            <category><![CDATA[business-intelligence]]></category>
            <category><![CDATA[math]]></category>
            <category><![CDATA[tableau]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Sat, 01 Aug 2026 05:01:03 GMT</pubDate>
            <atom:updated>2026-08-01T05:01:03.103Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*erzG9fa7h63DymkivPDQGg.png" /><figcaption>Thumbnail</figcaption></figure><p>I found this formula in a <a href="https://trailhead.salesforce.com/trailblazer-community/feed/0D5KX00000kRXMi0AO">Trailblazer Community thread</a> and it’s awesome. I modified it, and now reach for it any time a metric in Tableau needs to reformat itself based on its own size. I’ll lead with the calculation so you can bookmark it and copy/paste easily. Explanation to follow.</p><pre>STR(INT(ROUND(ABS(([Value]))/POWER(1000, INT(LOG(ROUND(ABS(([Value]))/10,0)*10,1000))),1)))<br>+<br>IF ABS([Value]) &lt; 1000<br>THEN &#39;&#39;<br>ELSEIF (ROUND(ABS(([Value]))/POWER(1000, INT(LOG(ABS(([Value])),1000))),1)*10)%10=0<br>THEN &#39;.&#39; + STR((ROUND(ABS(([Value]))/POWER(1000,INT(LOG(ABS(([Value])),1000))),1)*10)%10)<br>ELSE &#39;.&#39; + STR((ROUND(ABS(([Value]))/POWER(1000,INT(LOG(ABS(([Value])),1000))),1)*10)%10)<br>END<br>+<br>CASE<br>INT(LOG(ROUND(ABS(([Value]))/10,0)*10,1000))<br>WHEN 1 THEN &#39;K&#39;<br>WHEN 2 THEN &#39;M&#39;<br>WHEN 3 THEN &#39;B&#39;<br>WHEN 4 THEN &#39;T&#39;<br>ELSE &#39;&#39;<br>END</pre><blockquote>Pro tip: rename your metric to [Value], paste the calculation in, then rename the metric back. Tableau rewrites the references for you, and you never have to touch the formula.</blockquote><h3>The Problem</h3><p>Tableau’s built-in number formatting covers a lot of ground. You can format currency, set percentages, bolt on prefixes and suffixes, and give positives, negatives, and zeros their own treatment. One case it misses shows up in my work over and over: a BAN that switches between K, M, and B on its own, with flexible decimal precision.</p><p>These are the test cases I want a BAN to pass:</p><pre>    Value | Display<br>      545 | 545<br>    2,345 | 2.3K<br>    5,041 | 5.0K<br>  402,310 | 402.3K<br>9,145,232 | 9.1M<br>        2 | NULL</pre><p>The last case matters for real-world data. Values between 0 and 4 return blank instead of showing noise or test values.</p><p>You might reach for STR() and ROUND() to wrap our number and call it done. That breaks. Jonathan Drummey explains the reason in another <a href="https://trailhead.salesforce.com/trailblazer-community/feed/0D5KX00000kRSxE0AW">forum thread</a>: Tableau stores some values as floating point numbers that do not sit where you expect them to. 1.16 lives in memory as 1.1599999999…, so STR(ROUND(1.16, 2)) returns a mess.</p><p>The calculation above survives those cases. Here it is again with comments, so you can bend it to your own use case.</p><pre>// Digits left of the decimal (numbers less than 5 will round to 0)<br>STR(INT(ROUND(ABS(([Value]))/POWER(1000, INT(LOG(ROUND(ABS(([Value]))/10,0)*10,1000))),1)))<br>+<br>// Decimal and digits to the right of the decimal<br>// Case for 0-999 - no decimal<br>IF ABS([Value]) &lt; 1000<br>THEN &#39;&#39;<br>// Case when decimal value is 0 (Modify to select &quot;XXX.0M&quot; or &quot;XXXM&quot;)<br>ELSEIF (ROUND(ABS(([Value]))/POWER(1000, INT(LOG(ABS(([Value])),1000))),1)*10)%10=0<br>THEN &#39;.&#39; + STR((ROUND(ABS(([Value]))/POWER(1000,INT(LOG(ABS(([Value])),1000))),1)*10)%10)<br>// Case when decimal value is 1-9<br>ELSE &#39;.&#39; + STR((ROUND(ABS(([Value]))/POWER(1000,INT(LOG(ABS(([Value])),1000))),1)*10)%10)<br>END<br>+<br>// Determine the suffix (thousands/millions/billions/trillions)<br>CASE<br>INT(LOG(ROUND(ABS(([Value]))/10,0)*10,1000))<br>WHEN 1 THEN &#39;K&#39;<br>WHEN 2 THEN &#39;M&#39;<br>WHEN 3 THEN &#39;B&#39;<br>WHEN 4 THEN &#39;T&#39;<br>ELSE &#39;&#39;<br>END</pre><p>The ELSEIF and the ELSE branches intentionally look identical. The current version keeps the decimal for all rounded numbers, so you get 402.0K. Swap the ELSEIF result to “” and that same value drops to 402K.</p><h3>Only Read This Section If You’re a Math Nerd</h3><p>Instead of a wall of IF/ELSEIF branches covering every order of magnitude, this calc leans on base-1000 logarithms to work out which tier a number belongs to. Four nested pieces do that work.</p><h4>1. Finding the Tier with LOG</h4><pre>INT(LOG(ABS([Value]), 1000))</pre><p>LOG(number, 1000) answers one question: what power do you raise 1000 to in order to land on this number?</p><ul><li>500 gives 0.90</li><li>1,500 gives 1.06</li><li>2,500,000 gives 2.13</li></ul><p>Wrap that in INT(), lose the decimals, and you have a clean tier index.</p><ul><li>Tier 0: under 1,000</li><li>Tier 1: thousands</li><li>Tier 2: millions</li><li>Tier 3: billions</li></ul><p>Those indexes line up with the CASE statement at the bottom of the calc that hands out K, M, B, and T.</p><h4>2. Building the Divisor using POWER</h4><pre>POWER(1000, [Tier Index])</pre><p>Feed the tier index back in and you get the number you need to divide by. Index 1 returns 1,000. Index 2 returns 1,000,000. Divide [Value] by that result and the number shrinks to its display form, so 2,500,000 becomes 2.5.</p><h4>3. The Pre-Rounding Trick</h4><pre>ROUND(ABS([Value])/10, 0) * 10</pre><p>You’ll spot this tucked inside the LOG function in the first and third blocks. Take a value of 996. LOG(996, 1000) drops it into tier 0, so the calc prints 996 with no suffix at all. Dividing by 10, rounding to zero decimals, and multiplying by 10 pushes 996 up to 1000 before LOG ever sees it. Values from 995 to 999 land in tier 1 and pick up their K.</p><h4>4. Extracting the Decimal using Modulo (%)</h4><pre>(… * 10) % 10</pre><p>This one drives the middle block. Say your base number after the division step is 1.5. Multiply by 10 for 15, take 15 % 10, and you’re left with 5. The calc glues a decimal point in front of it and you get the .5 in 1.5K.</p><p>Though this is just a single calculation and a single tip, I use it so often that I wanted to dedicate a post to it. Give this calculation a try for your next Tableau project, you won’t regret it!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6908ac746836" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Bootstrap Sampling Can’t Invent Data You Never Collected]]></title>
            <link>https://medium.com/@feizadata/bootstrap-sampling-cant-invent-data-you-never-collected-604c111c1a6b?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/604c111c1a6b</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[statistics]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Wed, 01 Jul 2026 05:01:03 GMT</pubDate>
            <atom:updated>2026-07-01T05:01:03.386Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/720/1*ExhO3dPB6F9vUviySf8_9Q.png" /><figcaption>Thumbnail</figcaption></figure><p>There are some concepts from undergrad statistics that sit in the back of your mind for years. Bootstrap sampling was one of those for me.</p><p>The basic idea is this. When you collect a data sample in an attempt to estimate a parameter from a larger population (average, proportion, etc.), you should question how well the sample actually reflects reality. Bootstrap sampling is a technique for answering that. It uses the sample you already have to estimate how much your result would vary if you’d sampled slightly differently. It provides a confidence interval that captures where the true population value likely sits <strong>given the data you collected</strong>.</p><p>This can be an useful solution for a very common issue when sampling data. Only a few years after my introduction to bootstrap sampling, I mistakenly considered it for a real problem at work.</p><p>I’d been tasked with analyzing employee survey data across a handful of plant locations. By plant, survey responses from each employee were rolled up into a single NPS score to capture how they felt about work culture. I wanted to compare average scores across plants to understand differences in working environments and identify areas for improvement.</p><p>The analysis seemed simple until I looked at the final response rates by plant. Some plants had 80 responses, while others only had 8. Relative to the entire population size of each plant, low response rates would undermine any fair comparison between plants across the network.</p><p>That’s when I recalled bootstrap sampling. Perhaps I could bootstrap from a low-response sample to build confidence around my shallow estimate of a plant’s true average NPS?</p><h4>What Bootstrap Sampling Actually Does</h4><p>Say that 8 responses at a plant produced this average NPS score.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7d0UhqrH18s37gGNy14-Sw.png" /><figcaption>Each employee NPS score ranges from -100 to 100 points</figcaption></figure><p>Bootstrapping tells us we can draw a new sample of the same size (8) from the original sample with replacement. That means we pick a respondent from the original sample at random, collect their NPS score, and put them back in line. This creates a <strong>Bootstrap Sample. </strong>We can then repeat this process <em>many, many times</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9Qi8v-wKyhXXdlBnnsRjtg.png" /><figcaption>Repeating this process many (n) times…</figcaption></figure><p>Assume we resampled this data enough times to create <strong>n</strong> bootstrapped samples. For each resample, we can compute an average NPS score and plot that average on a shared histogram with the other resamples. We observe over time that, with enough resamples, a distribution begins to form around the original sample mean of 24.4.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*k0BiKvKxKRxmHlu-1LpPQQ.png" /></figure><p>The 95% confidence interval spans a range of 67 NPS points, telling us that with only 8 original responses, it’s difficult to pin down the true average NPS at this plant. The wide interval depicts uncertainty <em>within our original sample</em>.</p><p>If you take anything away from this post, let it be this: <strong>The distribution above says nothing about whether those 8 original responses were representative of the plant in the first place</strong>. Bootstrap sampling is great at explaining variance within the sample data we do have, but it can’t entirely make up for unknown data we failed to collect.</p><p>Imagine you wanted to understand how many people prefer heavy metal over classical music. You decide to survey 8 people leaving a metal concert. You could then bootstrap many resamples from those 8 responses to “confidently” claim that 85% to 100% of people prefer metal music. This confidence interval could look entirely reasonable, but the bootstrap had no way of knowing that those 8 responses were biased. The interval was a confident estimate of the wrong population.</p><h4>When Bootstrap Sampling Is the Right Tool</h4><p>Bootstrap sampling isn’t the villain here. The issue was trying to use it as a remedy for something it was never designed to fix. When you have a sample that’s large and reasonable enough to represent a population, the power of bootstrapping emerges. It places an honest confidence interval around an original sample estimate without forcing assumptions on the underlying data distribution.</p><p>It’s not uncommon for a statistical method to assume data follows a normal distribution. Real survey data (and especially NPS scores) tend to be polarized. Bootstrap sampling avoids this problem entirely by building a distribution <em>empirically </em>from the data you actually collected. It does not try to force the sample distribution into a shape that doesn’t fit.</p><p>Bootstrap sampling forms a distribution around a sample parameter estimate through repeated resampling. This distribution tells you how variable your estimate is given the data you sampled.<strong> It tells you nothing about the data you never collected.</strong></p><p>The mistake wasn’t in the tool. It was in thinking a statistical method could substitute for data I never had.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=604c111c1a6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Custom Keyboard Shortcuts in Tableau (or any Mac Application)]]></title>
            <link>https://medium.com/@feizadata/custom-keyboard-shortcuts-in-tableau-or-any-mac-application-9e96cd37c0b6?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/9e96cd37c0b6</guid>
            <category><![CDATA[keyboard-shortcuts]]></category>
            <category><![CDATA[datafam]]></category>
            <category><![CDATA[mac]]></category>
            <category><![CDATA[tableau]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 00:22:02 GMT</pubDate>
            <atom:updated>2026-06-02T00:22:02.356Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LF2TdZlhWkF-QFhDgjkY4A.png" /><figcaption>Thumbnail</figcaption></figure><p>When I switched from a Windows computer to a Mac, the loss of some keyboard shortcuts in Tableau dramatically impacted my productivity. I’d gotten used to many of the default Windows shortcuts like:</p><pre>─────────────────────────────────────────────────────────<br> Function                     Windows Keyboard Shortcut          <br>─────────────────────────────────────────────────────────<br> Create Calculated Field      Alt+A+C                    <br> Swap Rows and Columns        Ctrl+W                     <br> Export Packaged Workbook     Alt+F+K                    <br> Edit Tooltip                 Alt+W+O                    <br> Open Caption                 Alt+W+A+Enter              <br>─────────────────────────────────────────────────────────</pre><p>And when I switched to a Mac, there wasn’t a replacement keyboard command for these. Fortunately, there’s a way on Mac to create your own keyboard shortcuts for Tableau (and any other application on your machine).</p><p>First, navigate to <strong>System Settings &gt; Keyboard &gt; Keyboard Shortcuts &gt; App Shortcuts.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/685/1*J8Q8t9nKDKWqPMTjY1pClA@2x.png" /><figcaption>Window to Input Custom Mac Shortcuts</figcaption></figure><p>Click on the + icon and select the Application that you want to make a shortcut for (I’m using Tableau Desktop 2026.1).</p><p>Now, go through the options in your top Mac toolbar and identify any functions you’d like to create a shortcut for.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/801/1*jop6IFQiXNV4IrVBULt6sw.png" /><figcaption>Example of Functions to Convert to Shortcuts</figcaption></figure><p>Make sure you replicate these names one-to-one as they appear, and build your keyboard shortcuts. Here are my preferred shortcuts for Tableau:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/707/1*ff8x7IjdbcA2_DxOmsE1bg.png" /><figcaption>My Tableau Keyboard Shortcuts on Mac</figcaption></figure><p>I tend to use the “Command+Number” combination for functions I use regularly. Command+1 is actually more comfortable than Alt+A+C on a Windows machine now :)</p><p>Quick blog post this month, but this is a useful tip that I hope benefits some other Tableau developers on Mac devices. Thank you for reading!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9e96cd37c0b6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rethinking A/B Testing with Bayesian Inference]]></title>
            <link>https://medium.com/@feizadata/rethinking-a-b-testing-with-bayesian-inference-6f9dca94ed2f?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/6f9dca94ed2f</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[bayesian-statistics]]></category>
            <category><![CDATA[statistics]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Wed, 13 May 2026 05:01:00 GMT</pubDate>
            <atom:updated>2026-05-13T05:01:00.487Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X2x_jDDdyV8kGZKxD8KkYQ.png" /><figcaption>Thumbnail</figcaption></figure><p>Running an A/B test sounds straightforward. Split your test data, measure an outcome, and pick a winner. Yet assumptions in the most common A/B tests can break down in ways that are easy to miss and costly to ignore. Luckily, there are alternative approaches that treat belief as something to be updated as evidence arrives, rather than a mere threshold to be crossed. In this post, we’ll build our knowledge on the Bayesian statistical framework and examine what it really costs to get decisions wrong.</p><h3><strong>The Issue with Frequentist Probability</strong></h3><p>The <strong>Frequentist </strong>definition of probability is the most widely known in statistics. It states that probability measures the long-run frequency of an event occurring over many repeated trials. A classic example is the coin flip. Flipping a coin millions of times, the proportion of heads gradually stabilizes near 0.5 as an estimate of the true underlying probability. That true probability is always unknown, but with enough repeated trials, we can get reasonably close.</p><p><strong>Frequentist A/B Testing</strong> is an experiment we run to see if two processes behave significantly differently from one another. Consider a pharmaceutical company testing a new treatment to relieve headaches. Patients are randomly assigned to receive treatment A (an existing medication) or treatment B (the new medication). After exposure to the treatment, each patient reports whether their headache was relieved within one hour. The goal of this test is to determine whether treatment B produces a meaningfully higher relief rate than treatment A. The null hypothesis for this A/B test is: Treatments A and B produce identical relief rates. We’d then define a fixed number of patients, each receiving exactly one of the two treatments. Then, we’d use the test results to compute a single p-value answering the question: What’s the probability of observing data as extreme as ours if the null hypothesis is true? If this probability is very unlikely (&lt; 0.05), we’d reject the null hypothesis and claim that the two treatments produce meaningfully different relief rates, resulting in a winner.</p><p>Key takeaways on the Frequentist A/B approach:</p><ul><li>The p-value from this test is the probability of observing our collected data given the <strong>null hypothesis is true</strong>.</li><li>Frequentist probability treats the true relief rates of treatments A and B as <strong>fixed, unknown constants</strong> (probabilities that do not change).</li><li>The false positive rate (0.05) is only controlled <strong>if </strong>we commit to our sample size in advance and look at the results exactly once.</li></ul><p>Let’s say we decide a reasonable sample size for this experiment to be 500 patients. As we carry out this experiment, stakeholders behind the scenes are eager to monitor the results in real time. After the first 50 sessions, the resulting p-value dips to 0.03, indicating we should reject the null hypothesis and favor treatment B. The issue is, every time we glance at the accumulating results <strong>before </strong>reaching our defined sample size of 500, we risk rejecting the null hypothesis by chance alone. This is known as <strong>p-hacking by peeking</strong>, which can actually inflate the 0.05 false positive rate we thought we were controlling. Research suggests that peeking at live results just ten times can balloon a 0.05 false positive rate to upwards of 0.25 <a href="https://www.evanmiller.org/how-not-to-run-an-ab-test.html">(Miller, 2010)</a>. So why peek at all? Interim results can be tempting to those who have incentive to act on promising data early, even when the data isn’t fully conclusive. Frequentist A/B testing is meant for a world where we resist that pressure, but that’s not always the case in practice.</p><p>After 100 sessions, treatment B has a relief rate of 68% and treatment A has a relief rate of 61%. Is treatment B truly better yet? If we’re playing by the frequentist rules, we’re in a tough spot. We’ve been monitoring the number of collected sessions, but we still haven’t reached our pre-defined sample size of 500 (and frankly, we’re questioning whether 100 independent patient sessions could have been a sufficient sample size to begin with). At this point, the p-value still cannot tell us: <strong>Given the data, how confident should we be that treatment B is genuinely better?</strong> Lucky for us, Bayesian inference can help us answer this question.</p><p><strong>What is Bayesian Statistics?</strong></p><p>The <strong>Bayesian </strong>definition of probability is a measurement of belief. It represents a value of uncertainty about an event, whether it happens once or many times. Our belief in the outcome of an event can change as we observe new evidence. Instead of committing to a single fixed estimate, we maintain a full distribution of beliefs across all possible parameter values, updating these beliefs as new evidence arrives.</p><p>Let’s revisit the coin flip for a moment. Under frequentist thinking, we flip a coin millions of times, and the probability of landing on heads will converge to a fixed value around 0.5. With Bayesian probability, we start with a belief about the coin’s bias, like “I expect the true probability of landing on heads to fall between 0.4 and 0.6.” We might suspect the coin is fair, but we’re not certain. Well, we can revise that belief with every flip. After 10 flips, our distribution is wide and uncertain. After 10,000 flips, it has narrowed considerably. While both Frequentist and Bayesian estimates in this example might lead to the same result, the Bayesian approach allows us to carry a full picture of our estimates and uncertainty along the way.</p><p>We’ll use this same approach now for the clinical drug experiment. Instead of asking whether the relief rate of treatment B is significantly different than treatment A at some fixed sample size, we’ll construct a <strong>distribution of beliefs</strong> around both treatments’ true relief rates, updating both in real time as patient sessions are recorded. To understand how that updating works, we need Bayes’ Theorem. Let’s recall our favorite <strong>conditional probability</strong> formula:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/203/1*ZdRaIz5w1H012yf_UQlysg.png" /></figure><p>In our experiment, we’re not talking about abstract events A and B. We care about parameters and data. Let’s rewrite this formula as:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/203/1*qSZKuWGlcCl-dBCMiQjD8Q.png" /></figure><p>Where <strong>θ </strong>is the unknown parameter we want to learn (the true relief rate for a particular treatment) and <strong>X </strong>is the observed data (the recorded outcomes of relieved and not relieved from real patient sessions).</p><p>Next, we need to define a few new terms:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Aw193YoUsVcCbrizQl0GNQ.png" /></figure><p>The posterior is always what we’re after. It reflects our updated belief about θ given everything observed so far.</p><p>Let’s put this knowledge to the test with a famous example in Bayesian probability.</p><p><em>Consider a disease that affects 1% of the population. A diagnostic test for this disease is 95% accurate: if you have the disease, the test returns positive 95% of the time. If you don’t have the disease, the test returns negative 95% of the time (a 5% false positive rate). You test positive. What is the probability you actually have the disease?</em></p><p>Here’s how we go about answering this question using the formulas above.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/776/1*OQDVEOyLPDdh1zDYghNbEw.png" /></figure><p>Drumroll…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/392/1*wmEapWzd5-uHQdNR5f9YwQ.png" /></figure><p>There’s actually only a 16% chance you have the disease despite a 95% accurate test. Many people (my initial self included) assume the answer is closer to 95%. The disease is so rare that false positives outnumber true positives across the full population. Our prior belief (only 1% of people have this disease) is powerful enough to override a test that is 95% accurate. Thus, the posterior is a balance between prior belief and observed evidence. If you were to test positive twice in a row, that posterior of 16% becomes your new prior going into the next update. Each result reshapes your belief, which becomes the starting point for the next posterior.</p><p><strong><em>A side note on the likelihood P(X)</em></strong></p><p>Notice P(X) sits in the denominator of the posterior formula above. Its only job is to ensure the posterior sums to 1 across all possible values of θ. It has no dependence on θ and is a constant regardless of the parameter value we’re evaluating. Since we’ll be focused on identifying the <strong>shape of the posterior distribution</strong>, a constant that scales everything uniformly doesn’t meaningfully change anything. We can drop the denominator and write:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/210/1*0f2CljtCF_KXunm-104WmA.png" /><figcaption>Where ∝ means “<em>proportional to</em>”</figcaption></figure><p>Again, the shape of the posterior is determined <strong>entirely by the numerator</strong>, so we’ll refer to this formula from here on out.</p><p>Recall Frequentist probability views the true relief rate of treatment B as a fixed, unknown constant. We can only <strong>estimate </strong>this true relief rate by collecting data, running a test, and arriving at a single value like 68%. We can add a confidence interval around our estimate, but we’re ultimately left with a single estimate and a range. There’s no formal way for us to say we believe the true rate is probably around 68% with a reasonable chance it’s as high as 75%. We can’t update that estimate fluidly as new sessions arrive either. The Bayesian play is to represent θ (the true relief rate) as a <strong>probability distribution</strong>. Instead of committing to a single estimate, we capture a full distribution of our uncertainty across every possible value between 0 and 1. The mean of the distribution is where we think the true θ value most likely falls. The width of the distribution reflects how certain or uncertain we are. Whenever new data arrives and we update our prior, the entire distribution updates.</p><p>A <strong>Beta Distribution</strong> Beta(α, β<strong>)</strong> is a common probability distribution in Bayesian statistics defined on the interval [0,1]. The distribution<strong> </strong>uses two parameters, α (alpha) and β (beta)<strong>, </strong>to represent pseudo-counts of successes and failures measuring θ. We can use this distribution to model our prior belief of θ. For example, think of α as the number of relieved outcomes (successes) and β as the number of non-relieved outcomes (failures) before the clinical drug experiment begins. A Beta(1, 1) prior represents starting as if we’ve seen exactly one success and one failure (essentially no prior bias). A Beta(10, 30) represents having already observed 10 successes from 40 sessions, suggesting a prior belief of roughly 25% relief rate with moderate confidence.</p><p>If we have no prior beliefs on the true relief rates of treatments A or B at the start of the A/B test, we can use the neutral Beta(1,1) as the prior for both treatments. This is a flat, uninformative prior telling the model we have no strong belief about the effectiveness of either treatment before the experiment. After 500 sessions are collected for each treatment, each success and failure to relieve a headache would update the corresponding Beta posterior distributions (one for θ_A , one for θ_B). We’d then ask what the probability is that a sample drawn from B’s posterior exceeds a sample drawn from A’s posterior.</p><p>The probability density function for the Beta distribution is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/219/1*uB9i9yMdbGUyxEf980NUBQ.png" /></figure><p>Where B(α, β) is the normalizing constant ensuring the distribution integrates to 1 over [0,1]. This is analogous to P(X) in Bayes’ theorem. When we plot the Beta distribution, θ sits on the x-axis. We’re never claiming we know θ. Instead, we’re asking how likely each possible value of θ is given our current beliefs. We’ll take a look at a few of these plots below. Try thinking of them as pictures of uncertainty around the true value of θ.</p><p>Here are some examples of a prior Beta distribution for the relief rate of treatment B. The peak location is always α/(α+β). The width tells us how much volatility is behind that belief.</p><p><strong>Beta(1, 1) — An uninformative prior</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/953/1*TWk6eo_1c9ZaYUi4bN1URg.png" /></figure><p>This is a completely flat distribution. Every possible relief rate from 0 to 1 is equally likely. We’re essentially expressing no prior belief about what to expect.</p><p><strong>Beta(13, 7) — A weak prior belief</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/1*OfS76pn3kfoug4f-C_6XBA.png" /></figure><p>Equivalent to observing 13 relieved outcomes from 20 sessions. Soft belief that treatment B’s relief rate is around 65%.</p><p><strong>Beta(204, 96) — A strong prior belief</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/970/1*EuTTUnn3R5BDBPXggp_2Kw.png" /></figure><p>Equivalent to 204 relieved outcomes from 300 sessions. High confidence that treatment B’s relief rate is near 68%.</p><p>Now, here’s where the math gets exciting. If we pair the right prior P(θ) distribution with the right likelihood P(X∣θ) distribution, the posterior comes out as the same family of distributions as the prior. This is called C<strong>onjugacy</strong>, and it’s what makes our update rule clean enough to derive by hand. In the clinical drug experiment, our likelihood comes from a binomial distribution (relieved or not relieved being the binary outcome), and the prior comes from a Beta distribution (defined on the closed [0,1] interval for probability). Together, their combination forms a Beta-Binomial conjugate pair, which can be used to update the prior.</p><p>Let’s break down how this works. Consider the likelihood P(X|θ). If we observe n patient sessions for a particular treatment with k successful relief outcomes, the likelihood of this data given a particular value of θ is:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/226/1*p0CmxIl8BPiPvcsJVYUQsw.png" /></figure><p>This is the Binomial distribution. It answers the question of how probable it is to observe exactly k relief outcomes from n sessions, given a true relief rate of θ. Notice that (n choose k), the binomial coefficient, is only a constant with no dependence on θ. Just like P(X) in Bayes’ theorem, we can drop it when reasoning about the shape of the posterior. So, the likelihood simplifies to:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/188/1*pb7fZnOX8ADNkBPz5srZfQ.png" /></figure><p>To determine the Posterior P(θ|X), we must multiply the likelihood above with our Beta prior. Recall this prior takes the form:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/173/1*_R0cFuiY7v7PCcb1c-3IvQ.png" /></figure><p>Multiplying the likelihood by the prior:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/315/1*GYUutUOYPIKYqMaW-7j3vA.png" /></figure><p>Collecting exponents:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/257/1*HUgWMt5QbuHQ7jDywSdKVw.png" /></figure><p>Upon simplifying, this looks exactly like our Beta distribution, just with updated parameters:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/274/1*SUG2Qna0sgM4jX8utH1Q2Q.png" /></figure><p>That’s the entire update rule! For a specific treatment in the experiment, our prior was B(α, β). We observe k relieved outcomes from n sessions for that treatment. So, our final posterior is B(α+k, β+n-k).</p><p>Here’s an example of how our belief in the true relief rate of treatment B can change as we accumulate data and update our prior with the latest posterior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*34wHQ-syiMYB9AdURDUq3w.png" /></figure><p>Notice that the mean of the posterior after each data collection converges to the true relief rate as we accumulate more and more patient sessions. Each row’s posterior becomes the prior for the next update. Notice also that the first prior’s influence pulling the mean towards 0.5 fades away quickly as real data accumulates. By session 500, the data has taken over to shape the mean. In the clinical drug experiment, we can run this update simultaneously for both treatments, maintaining Beta(α_A, β_A) and Beta(α_B, β_B) in parallel. Every session that arrives updates one of the two distributions depending on which treatment was prescribed. We are never forced to wait for a pre-defined sample size, and the posteriors simply reflect whatever evidence we’ve seen so far.</p><h3><strong>Making Decisions Under Uncertainty</strong></h3><p>At any point in the clinical drug trial, we’ll have two posterior distributions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/234/1*t8kj00l1ryx5NMW4DvG-EA.png" /></figure><p>Either distribution represents our full belief about that treatment’s true relief rate at that point in the experiment. However, these distributions aren’t enough to guide decisions yet. When a stakeholder asks if treatment B is truly better than treatment A, we must answer two questions first.</p><ol><li>Given the latest data, how likely is it that treatment B is truly better than A?</li><li>What would a wrong decision in the experiment cost us?</li></ol><p>To answer question 1…</p><p>A<strong> Frequentist 95% confidence interval</strong> built around an estimated relief rate won’t tell us there’s a 95% probability the true rate lives inside it. It means that across many repeated experiments, 95% of the constructed intervals would contain the true rate. The true relief rate is always fixed in this lens. A specific confidence interval has it or it doesn’t, and frequentist statistics doesn’t allow us to assign a probability to that.</p><p>A <strong>Bayesian 95% credible interval</strong> is exactly what it sounds like. Given the data observed, there is a 95% probability that the true relief rate lies within this range. In the clinical drug trial, by the time treatment B’s posterior is Beta(341,161), we can hand a stakeholder a credible interval and say with confidence exactly where the true relief rate most likely falls.</p><p>With two independent Beta posterior distributions, we want to know how often a sample drawn from B’s posterior would exceed a sample drawn from A’s posterior. This can be denoted:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/89/1*GPVW3CaBlaDSoURU05O6cw.png" /></figure><p>Analytically, this is computed by integrating over all possible values of θ_A​ and θ_B​:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/321/1*uzQmq_NhGeSdK2dZLFuXiA.png" /></figure><p>The double integral may look intimidating, but the intuition is straightforward. If treatment B’s posterior is almost entirely to the right of treatment A’s on the relief rate axis, P(B&gt;A) approaches 1. If the two distributions heavily overlap, P(B&gt;A) hovers near 0.5, meaning the data hasn’t separated them yet. P(B&gt;A) is useful, but it is not quite the right number to make a final decision on.</p><p>To answer question 2…</p><p>P(B&gt;A) = 0.95 might sound compelling, but what if treatment B is only 0.1% better than treatment A in the scenarios where it wins and 5% worse in the scenarios where it loses? Calling treatment B better might still be the wrong call. This is where expected loss comes in. What is the average cost of making the wrong decision?</p><p>We can define a<strong> loss function</strong> for each possible decision (choosing treatment A or B as superior). In either case, the loss measures how much effectiveness we leave on the table by promoting the inferior treatment. If we promote treatment B and it turns out to be worse, the loss is how much A exceeded B. If we promote treatment A and it turns out to be better, the loss evaluates to zero.</p><p>If we were to promote treatment B:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/276/1*ewkYft2e1DZYDcH_GnAbUQ.png" /></figure><p>Similarly, if we were to promote treatment A:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/282/1*rTE3ufayNFhmlLoKRcUrjQ.png" /></figure><p>We would only decide to promote Treatment B if the expected loss falls below some acceptable threshold ε:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/142/1*37D0gUhUdLSw5wbRUD_D-A.png" /></figure><p><em>For simplicity, we use a shared threshold ε since treatment A is already the incumbent treatment. We could technically introduce a separate ε for promoting an alternative depending on the experiment.</em></p><p>Remember that <strong>Frequentist </strong>A/B testing should never allow early stopping in an experiment. The moment we break our commitment to the sample size, our false positive rate inflates and the results become unreliable. Under a <strong>Bayesian </strong>A/B test, we can stop as soon as the expected loss drops below ε. Whether it takes 100 or 1,000 sessions in the clinical drug experiment, the data will tell us once we’ve collected enough of it.</p><h3>Tying It All Together</h3><p>Let’s close with one last scenario. Imagine you’re a data scientist working for a video streaming platform. The platform recommends new content to viewers as soon as they finish watching a video. You’ve built a new recommendation model and want to know if it drives more clicks than the existing one.</p><p>Model A is the existing model. It recommends content based on what similar viewers have watched. Model B is your new model. It factors in how long the viewer watched their last video for and whether they’ve rewatched content before. Both models show a ranked list of recommendations at the end of every video. The metric is simple enough. Did the viewer click on at least one recommended video?</p><p>We can carry out a simulation exposing 500 viewer sessions to each model. The key unknown in this experiment is the click-through rate (CTR) generated by each model’s recommendation. For our purposes, we’ll set the ground truth CTR for Models A and B at 25% and 28% respectively. This is a modest lift, but reasonable in practice. A 3% difference is meaningful at scale, but subtle enough to go undetected in the early stages of an experiment. That’s what will make this an interesting test case.</p><p>Let’s fire up our Python IDE to get started. We’ll first need to simulate viewer session outcomes from the ground truth Binomial distributions of each model.</p><pre># Import libraries<br>import numpy as np<br>from scipy import stats<br>import matplotlib.pyplot as plt<br>from scipy.special import betaln<br>from scipy.stats import chi2_contingency<br><br># Set seed<br>np.random.seed(100)<br><br># Pre-defined click through rates for simulation (unknown during experiments)<br>true_ctr_a = 0.25<br>true_ctr_b = 0.28<br><br># Number of viewer sessions exposed to each model<br>n_sessions = 500<br><br># Simulate binary outcomes for each model<br>clicks_a = np.random.binomial(1, true_ctr_a, n_sessions)<br>clicks_b = np.random.binomial(1, true_ctr_b, n_sessions)<br><br># Print results<br>print(f&quot;Model A: {clicks_a.sum()} clicks from {n_sessions} sessions &quot;<br>      f&quot;(observed CTR: {clicks_a.mean():.3f})&quot;)<br>      <br>print(f&quot;Model B: {clicks_b.sum()} clicks from {n_sessions} sessions &quot;<br>      f&quot;(observed CTR: {clicks_b.mean():.3f})&quot;)</pre><p>Model A: 124 clicks from 500 sessions (observed CTR: 0.248)</p><p>Model B: 132 clicks from 500 sessions (observed CTR: 0.264)</p><p>We’ll now define a<em> </em>few functions to call later in the script. These pull from everything we’ve covered about Bayesian inference so far.</p><pre># Create a function to update priors from observed data<br>def update_posterior(alpha, beta, click):<br>    &quot;&quot;&quot;<br>    Update a Beta posterior given a single binary outcome.<br>    click = 1 (clicked) increments alpha (successes)<br>    click = 0 (no click) increments beta (failures)<br>    &quot;&quot;&quot;<br>    alpha += click<br>    beta += 1 - click<br>    return alpha, beta<br><br># Create a function to calculate the mean of a posterior distribution<br>def posterior_mean(alpha, beta):<br>    return alpha / (alpha + beta)<br><br># Create a function to calculate the variance of a posterior distribution<br>def posterior_variance(alpha, beta):<br>    return (alpha * beta) / ((alpha + beta) ** 2 * (alpha + beta + 1))<br><br># Create a function to return a credible interval containing 95% of the posterior distribution&#39;s belief about theta<br>def credible_interval(alpha, beta, credibility=0.95):<br>    lower = (1 - credibility) / 2<br>    upper = 1 - lower<br>    return stats.beta.ppf([lower, upper], alpha, beta)<br><br># Create a function to sample n theta values from a Beta posterior distribution<br>def sample_posterior(alpha, beta, n=100_000):<br>    return np.random.beta(alpha, beta, n)<br><br># Create a function to compute P(theta_B &gt; theta_A) for two Beta distributions<br>def prob_b_beats_a(alpha_a, beta_a, alpha_b, beta_b):<br>    total = 0.0<br>    &quot;&quot;&quot;<br>    Sum over each possible success count in B&#39;s posterior<br>    Given i successes for B, how much of A&#39;s distribution falls below that probability?<br>    betaln computes log Beta function values to keep numbers stable at large alpha/betas<br>    &quot;&quot;&quot;<br>    for i in range(alpha_b):<br>        total += np.exp(<br>            betaln(alpha_a + i, beta_a + beta_b) -<br>            np.log(beta_b + i) -<br>            betaln(1 + i, beta_b) -<br>            betaln(alpha_a, beta_a)<br>        )<br>    return total</pre><p>Unlike the last P(B &gt; A) function above, the <strong>expected loss integral</strong> doesn’t have a clean closed-form solution. We’ll use Monte Carlo sampling to draw a large number of θ (CTR) values from both posteriors, comparing them element-wise to simulate the joint distribution of (θ_A, θ_B) and estimate expected loss directly. The law of large numbers guarantees this estimate converges to the true expectation quickly. 100,000 samples will be plenty!</p><pre>def expected_loss(alpha_a, beta_a, alpha_b, beta_b, n_samples=100_000):<br>    &quot;&quot;&quot;<br>    Estimate expected loss for promoting B and promoting A<br>    via Monte Carlo sampling over the joint posterior.<br>    &quot;&quot;&quot;<br>    samples_a = sample_posterior(alpha_a, beta_a, n_samples)<br>    samples_b = sample_posterior(alpha_b, beta_b, n_samples)<br><br>    # Cost of promoting B: how much does A exceed B when A is better?<br>    loss_promote_b = np.mean(np.maximum(samples_a - samples_b, 0))<br><br>    # Cost of promoting A: how much does B exceed A when B is better?<br>    loss_promote_a = np.mean(np.maximum(samples_b - samples_a, 0))<br><br>    return loss_promote_b, loss_promote_a</pre><p>The result of this expected loss function is only useful when paired with a threshold. Recall from earlier that we define an acceptable threshold ε and only promote Model B if the expected loss of shipping Model B is less than ε. We’ll set ε equal to 0.5%, meaning we’re willing to accept a 0.5% hit to our clickthrough rate if we accidentally promote the wrong model.</p><pre># Maximum acceptable CTR loss<br>epsilon = 0.005</pre><p>Assuming we have no prior beliefs about the effectiveness of Model B, we’ll initialize its prior distribution as Beta(1,1). For Model A, its prior performance in production will help construct our prior belief. Let’s represent this belief in Model A’s CTR with a Beta(1,3) distribution.</p><pre># Initialize Beta(1,3) prior for Model A (estimating a 25% CTR)<br>alpha_a, beta_a = 1, 3<br><br># Initialize Beta(1,1) prior for Model B (flat, unbiased)<br>alpha_b, beta_b = 1, 1</pre><p>Next, let’s create a dictionary that logs the evolution of each posterior distribution as data accumulates.</p><pre># Track experiment history session by session<br>history = {<br>    &quot;sessions&quot;: [],<br>    &quot;mean_a&quot;: [],<br>    &quot;mean_b&quot;: [],<br>    &quot;prob_b_beats_a&quot;: [],<br>    &quot;loss_promote_b&quot;: [],<br>    &quot;loss_promote_a&quot;: [],<br>    &quot;decision_made&quot;: None<br>}<br><br># Record initial priors before any data arrives<br>history[&quot;sessions&quot;].append(0)<br>history[&quot;mean_a&quot;].append(posterior_mean(alpha_a, beta_a))<br>history[&quot;mean_b&quot;].append(posterior_mean(alpha_b, beta_b))<br>history[&quot;prob_b_beats_a&quot;].append(prob_b_beats_a(alpha_a, beta_a, alpha_b, beta_b))<br>history[&quot;loss_promote_b&quot;].append(0.0)<br>history[&quot;loss_promote_a&quot;].append(0.0)</pre><p><em>As you and I know at this point, Model B is the superior model, so we’ve shaped some of our functions and simulation results around capturing when we’ve confidently observed this outcome.</em></p><p>Let’s run our first simulation to see when Model B proves victorious in the A/B test!</p><pre>for t in range(1, n_sessions + 1):<br><br>    # Update each posterior with the current session outcome<br>    alpha_a, beta_a = update_posterior(alpha_a, beta_a, clicks_a[t-1])<br>    alpha_b, beta_b = update_posterior(alpha_b, beta_b, clicks_b[t-1])<br><br>    # Compute P(B &gt; A) and expected loss at this point in the experiment<br>    p_b_wins = prob_b_beats_a(alpha_a, beta_a, alpha_b, beta_b)<br>    loss_b, loss_a = expected_loss(alpha_a, beta_a, alpha_b, beta_b)<br><br>    # Record metrics for the current session<br>    history[&quot;sessions&quot;].append(t)<br>    history[&quot;mean_a&quot;].append(posterior_mean(alpha_a, beta_a))<br>    history[&quot;mean_b&quot;].append(posterior_mean(alpha_b, beta_b))<br>    history[&quot;prob_b_beats_a&quot;].append(p_b_wins)<br>    history[&quot;loss_promote_b&quot;].append(loss_b)<br>    history[&quot;loss_promote_a&quot;].append(loss_a)<br><br>    # Promote Model B when expected loss clears epsilon<br>    if history[&quot;decision_made&quot;] is None and loss_b &lt; epsilon:<br>        history[&quot;decision_made&quot;] = t<br>        print(f&quot;Decision at session {t}: Promote Model B&quot;)<br>        print(f&quot;P(B &gt; A):          {p_b_wins:.4f}&quot;)<br>        print(f&quot;Expected loss:     {loss_b:.5f}&quot;)<br>        print(f&quot;Posterior mean A:  {posterior_mean(alpha_a, beta_a):.4f}&quot;)<br>        print(f&quot;Posterior mean B:  {posterior_mean(alpha_b, beta_b):.4f}&quot;)</pre><p>Decision at session <strong>169</strong>: Promote Model B</p><p>P(B &gt; A): <strong>0.8267</strong></p><p>Expected loss: <strong>0.00429</strong></p><p>Posterior mean A: <strong>0.2254</strong></p><p>Posterior mean B: <strong>0.2690</strong></p><p>By session <strong>169, </strong>the expected loss drops below ε, triggering an early decision in favor of Model B. As you can see above, the posterior means didn’t converge to their true CTR values yet. We’re only stating the scenarios sampled from both posteriors produced an acceptably low cost of being wrong about promoting Model B. The Bayesian framework didn’t wait for certainty around the CTR estimates. Instead, the test stopped when the risk of a wrong decision became tolerable.</p><p>To put the Bayesian approach in context, let’s run the same simulation as a Frequentist A/B test. We’ll use a Chi-square test to evaluate results at exactly 169 sessions, with a fixed sample size of 500 viewer sessions per model. Stopping here (peeking early) will highlight just how differently the two frameworks interpret the same underlying data.</p><pre># Run frequentist chi-square test on data up to the Bayesian decision point<br>t_decision = history[&quot;decision_made&quot;]<br><br># Build a contingency table to run the chi-square test on<br>contingency = np.array([<br>    [clicks_a[:t_decision].sum(), t_decision - clicks_a[:t_decision].sum()],<br>    [clicks_b[:t_decision].sum(), t_decision - clicks_b[:t_decision].sum()]<br>])<br><br># Results<br>chi2, p_value, _, _ = chi2_contingency(contingency)<br>print(f&quot;Frequentist test at session {t_decision} (Peeking Early):&quot;)<br>print(f&quot;Chi-square statistic: {chi2:.4f}&quot;)<br>print(f&quot;p-value:              {p_value:.4f}&quot;)<br>print(f&quot;Significant at 0.05:  {p_value &lt; 0.05}&quot;)</pre><p>Frequentist test at session <strong>169 </strong>(Peeking Early):</p><p>Chi-square statistic: <strong>0.5749</strong></p><p>p-value: <strong>0.4483</strong></p><p>Significant at 0.05: <strong>False</strong></p><p>By session 169 in the Frequentist test, a p-value of 0.45 is nowhere near a 0.05 threshold. The test can’t reject the null hypothesis at this point. It demands more data before drawing any meaningful conclusion.</p><p>The Bayesian framework reached a decision at that same session without violating a single assumption. It’s worth noting again: P(B &gt; A) never crossed 0.95 across the entire experiment. The decision wasn’t driven by confidence in the direction of the result. It came down to the expected cost of promoting the wrong model falling below what we were willing to accept. Here’s how that played out across all 500 sessions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JXhDSHUZaQh1n1hqH_1fgA.png" /></figure><p>Plot 1 shows both posterior means converging toward the true CTR values over time. Increasing the number of viewer sessions from 500 to 1,000 would get us even closer to 0.25 and 0.28.</p><p>In plot 2, P(B &gt; A) spends a long stretch below 0.5 early on. The data briefly favored Model A before enough evidence accumulated to separate the two models. You’ll notice it never crosses 0.95, highlighting why expected loss is the better decision criterion.</p><p>Plot 3 tells the clearest story. Expected loss starts high and noisy, then steadily falls as the posteriors sharpen. It crosses epsilon at session 169 and lands us at our answer to promote Model B.</p><p>Bayesian statistics gives us a principled way to make decisions under uncertainty without the constraints of frequentist testing. We maintain a full distribution of beliefs about unknown parameters and update them continuously as new evidence arrives. This lets us answer questions that actually matter in practice. The expected loss framework takes it one step further, replacing an arbitrary significance threshold with a concrete measure of what being wrong actually costs. Next time you’re tempted to peek at that p-value, maybe just let the posterior do the talking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6f9dca94ed2f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Best Way to Dissect a Tableau Workbook]]></title>
            <link>https://medium.com/@feizadata/the-best-way-to-dissect-a-tableau-workbook-1388f9ae42c9?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/1388f9ae42c9</guid>
            <category><![CDATA[tableau]]></category>
            <category><![CDATA[data-lineage]]></category>
            <category><![CDATA[datafam]]></category>
            <category><![CDATA[tableau-desktop]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Sun, 01 Mar 2026 06:01:00 GMT</pubDate>
            <atom:updated>2026-03-01T06:01:00.746Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*32c60CL0oa4uGqQnHsd51g.png" /></figure><p>Have you ever inherited a workbook that needs updating and found yourself swimming in hundreds of calculated fields with no clear direction? Or maybe it’s your own old workbook that you’re revisiting, and you have absolutely no interest in dissecting it because it has grown into a monster.</p><p>Fortunately, about a year ago, Ana Rey de Castro built a solution for exactly this problem with her Tableau Calculation Extractor. The code is available for free on GitHub and runs locally on your machine, so there is no risk to your files. If you want a walkthrough of the tool, you can also check out a video from Tableau Tim in October 2024 where he discusses the release.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FfgJjpCZ5Rxc%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DfgJjpCZ5Rxc&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FfgJjpCZ5Rxc%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/aeff2a57da367c82082a3a36cc1a09a1/href">https://medium.com/media/aeff2a57da367c82082a3a36cc1a09a1/href</a></iframe><p>I first heard about this tool directly from Ana and used it shortly after to work through a few messy .twbx files. It was incredibly helpful at the time.</p><p>Recently, I ran into another use case for the tool, but this time I was working on a Mac instead of a Windows machine. One limitation of Ana’s original version was that a few of the required packages were Windows-specific, which made the code unusable on macOS.</p><p>With the help of Claude Code, I was able to port Ana’s project over to Mac and successfully run the tool locally on my machine. I’m excited to walk through that process and share it with you today.</p><h3>Where Can I Find the Code?</h3><p>You can find both Ana’s Windows version and the new Mac version here:</p><p><strong>Windows:</strong> <a href="https://github.com/scinana/tableauCalculationExport">https://github.com/scinana/tableauCalculationExport</a></p><p><strong>Mac:</strong> <a href="https://github.com/blakefeiza/Tableau-Calculation-Extractor-Mac/tree/main">https://github.com/blakefeiza/Tableau-Calculation-Extractor-Mac</a></p><h3>How Does it Work?</h3><p>The first step is to download the code to your local machine and open the project directory in your IDE of choice. For this walkthrough, I’ll be using Visual Studio Code, but any editor will work just fine.</p><blockquote>Since I’m on a Mac, this writeup reflects a macOS setup. If you’re working on Windows, I recommend checking out the Tableau Tim video mentioned above, along with Ana’s original walkthrough.</blockquote><p>After downloading the repository, create two folders in the root directory: one called inputs and one called outputs. The inputs folder is where you’ll place the .twbx file you want to analyze. The outputs folder is where the generated PDF, XLSX, or HTML files will appear after running the notebook.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zvqYeSjHi0vjLQwKDBBtZQ.png" /></figure><p>For this demo, I’ll be using my <a href="https://public.tableau.com/app/profile/blake.feiza/viz/365DaysofPlay/RadialSportsCalendar"><strong>365 Days of Play</strong></a> workbook. I downloaded the packaged workbook from Tableau Public and dropped the .twbx file into the inputs folder we just created.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7Gz4XvBrFzEuPC5aa1RLGA.png" /></figure><p>Now we’re ready to open the notebook.</p><p>In the top cell, review the required packages and confirm that everything is installed on your machine. I left the most common dependencies listed in that first cell for convenience. If anything is missing, go ahead and install it before running the rest of the notebook.</p><p>Once the dependencies are handled, the notebook should run in just a few seconds. After it completes, check your outputs folder. You should see the generated artifacts waiting for you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6rvLoHCv3qAcaEFv_qbSDQ.png" /></figure><p>Both the PDF and Excel outputs provide a comprehensive table listing all field names, field types, calculation logic, and associated data sources for the entire workbook.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VR8_uVLaYPAYNT0T95ox4g.png" /></figure><p>One difference you may notice on the Mac version is that instead of the static Mermaid diagram used in the Windows version, this implementation renders a dynamic Cytoscape diagram. The result is an interactive view of your calculation lineage.</p><p>I found this especially helpful when working with larger workbooks. You can drill into individual calculations, isolate the lineage for a specific field, and quickly collapse or expand sections as needed. You can also drag and reposition the nodes to structure the diagram in whatever way makes the most sense to you.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*wnxi4asZjVQUT7I_fUm-5Q.gif" /></figure><p>This was a shorter post, but I hope some of you find value in this community-built tool. If it makes your life a little easier the next time you need to retrace steps in a complex work project (or a Workout Wednesday challenge), then it has done its job.</p><p>This is more of a proof of concept than a fully polished application. If you have ideas on how to improve it, I’d love to hear them. And once again, a huge thank you to Ana Rey de Castro for building the original version and sharing it with the Tableau Community!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1388f9ae42c9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Overthinking Trellis Charts: Two Calculations]]></title>
            <link>https://medium.com/@feizadata/stop-overthinking-trellis-charts-two-calculations-4233b4dca912?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/4233b4dca912</guid>
            <category><![CDATA[datafam]]></category>
            <category><![CDATA[tableau]]></category>
            <category><![CDATA[table-calculation]]></category>
            <category><![CDATA[charts]]></category>
            <category><![CDATA[trellis]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 01:36:23 GMT</pubDate>
            <atom:updated>2026-07-31T21:35:10.982Z</atom:updated>
            <content:encoded><![CDATA[<pre>// Cols<br>(INDEX()-1)%[pCols]<br><br>// Rows<br>INT((INDEX()-1)/[pCols])</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6uXU7O_Yx3hCHCBzbpMqPA.png" /><figcaption>Thumbnail.</figcaption></figure><p>If you’ve spent some time scrolling on Tableau Public or observed enough data visualizations across the broader internet, odds are you’ve run into a <strong>Trellis Chart</strong>. Trellis Charts are useful because they segment a visual into small, repeated views. These views often share the same scales, and make patterns or differences easy to spot. They let viewers compare categories side-by-side without requiring much brain power.</p><h3>Where Does the Name “Trellis” Come From?</h3><p>Trellis Charts (also sometimes referred to as small multiples, lattice or panel charts) arrange the views in a grid-like layout. When built tall, these charts resemble a garden trellis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*-NrNRzhVKJgttIcQk2Vp8w.png" /><figcaption>Photo from <a href="https://www.wayfair.com/sca/ideas-and-advice/guides/what-is-a-trellis-T5945">Wayfair</a>, displaying a garden trellis.</figcaption></figure><h3>Building a Trellis Chart</h3><p>For this blog post, we will visualize the trend of wins per season for every NFL team. The data used for this analysis comes from the GitHub repo <a href="https://github.com/nflverse/nflreadpy">nflreadpy</a>. Credit to my brother Ethan Feiza for curating the dataset used in this analysis. You can find a copy of the final dataset <a href="https://data.world/blakefeiza/feizadata-february-2025-nfl-wins-coaches">here</a>.</p><p>While we could certainly plot all of these teams on a single-axis line chart and use a highlight action/color palette to select one or many teams to compare, that approach requires either a pre-baked insight or an end user’s appetite to explore the visualization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/843/1*Pp6Igo1ayOs-QkkiSWJoFA@2x.png" /><figcaption>Grey bowl of spaghetti slapped on a time series axis.</figcaption></figure><p>If we use a Trellis Chart instead, we can enable the user to make their own comparisons much easier with a static image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9MpgoQl4e7rT1RRVsA38qw@2x.png" /><figcaption>Final Trellis Chart from this blog post.</figcaption></figure><p>In order to create this chart, there are <strong>only two primary calculations</strong> to split up the view, one called <strong>“Rows”</strong> and one called <strong>“Cols”</strong>. After we walk through the initial two calculations, I’ll cover a few of the map layers tricks I chose to add on this viz but note that they are optional additions.</p><p>I’ll start by adding our <strong>[Season] to Columns</strong>, <strong>[Wins] to Rows</strong>, <strong>[Team] to Detail</strong>, and change the <strong>mark type to Area</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B7yWKKDidcjHe3qZesNT6Q@2x.png" /><figcaption>Default Tableau view after adding a few pills to the area chart.</figcaption></figure><p>Now in order to turn this mess into a trellis chart, we need to add two calculated fields that are each tied to a parameter I like to call <strong>[pCols]</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/553/1*EoI-Gvb3goLPF4wP3tRT_A@2x.png" /><figcaption>Parameter Creation Config Menu.</figcaption></figure><p>Once you’ve created this parameter, create the following two calculated fields:</p><pre>// Cols<br>(INDEX()-1)%[pCols]<br><br>// Rows<br>INT((INDEX()-1)/[pCols])</pre><p>Using some modular arithmetic and rounding, these calculations and our parameter give us the ability to specify how many columns we’d like in our trellis chart and the rows will update dynamically.</p><p>Add <strong>[Cols] to the Columns</strong> and <strong>[Rows] to the Rows</strong>. You should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EiQCsVGCbJHvRdqnFD1oWQ@2x.png" /><figcaption>Default Tableau view after adding [Rows] and [Cols] to their shelves.</figcaption></figure><p>The next thing we need to do is <strong>configure our table calculations</strong> for [Cols] and [Rows]. It doesn’t matter which one we start with because <strong>the steps are identical for both</strong>.</p><p>Right click the blue pill and select “Edit Table Calculation”. Select “Specific Dimensions”, make sure <strong>both your [Season] and [Team] are checked</strong>. <strong>Drag the [Team] field to the top</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/341/1*1bz-nc3Ga0uiNyzoWgIcDQ@2x.png" /><figcaption>Configuring Table Calculation for [Rows] and [Cols] (no map layers)</figcaption></figure><p>By default, our sort order will be alphabetical. If we wanted to do something like order the teams by total wins, we can configure that in the “Sort order” option as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/517/1*jETWMevPVRrly0qS36APZA@2x.png" /><figcaption>Updating the sort logic for [Cols] and [Rows] in the “Edit Table Calculation” box.</figcaption></figure><p>Follow these steps for both [Cols] and [Rows] and you should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*rT_gi2rauokFa7qX3CQiTQ@2x.png" /><figcaption>Tableau view after editing table calculations (no map layers).</figcaption></figure><p>From here, we can <strong>uncheck “Show Header”</strong> for our [Rows] and [Cols] fields (and optionally clean up [Season] and [Wins]), add some color or additional formatting clean-up, and get something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/986/1*15hWctVuITd4UZrTmpSzuA@2x.png" /><figcaption>Final chart option without map layers.</figcaption></figure><p>There are some traditional approaches to add labels and additional formatting to our chart from here, but I’ll leave that as an exercise to the reader. Instead, I’d like to dive into the “Map Layers alternate solution” instead to do as much custom formatting as our hearts desire :)</p><h3>Optional: Using Map Layers with Trellis Chart Calculations</h3><p>Okay, I know I said we’d only use two calculations, but this is the bonus section. There are a few more calculations that we’ll need to make in order to turn the above chart into a grid of maps. If you’d like to learn more about these calculations, the in-line comments will prove useful.</p><pre>// Normalized Season<br>// Using the traditional &quot;(Number - Min) / (Max - Min) normalization<br>// This gives us a value between 0 and 1 for every season.<br>// For a map, we need to stay between (-180,180). Staying between 0 and 1 makes the math easier later on.<br>([Season] - {MIN([Season])})<br>/<br>({MAX([Season])} - {MIN([Season])})<br><br>// Wins per Season Makepoint<br>// This plots a point on our map at <br>// Remember MAKEPOINT() uses (Latitude, Longitude) -&gt; backwards from (X,Y)<br>// I like to think of it as MAKEPOINT(Rise, Run)<br>MAKEPOINT(SUM([Wins]), MIN([Normalized Season]))</pre><p>On a new sheet, <strong>double click [Wins per Season Makepoint]</strong>, change your <strong>Mark Type to Line</strong>, drag <strong>[Team] to Detail</strong>, and <strong>[Season] to Path</strong>. You should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TTJ2y85efy3y2SNDc7S4KA@2x.png" /><figcaption>Default Tableau view on a map after adding some fields.</figcaption></figure><p>Next, add the fields <strong>[Rows] to Rows</strong> and <strong>[Cols] to Columns</strong>. This time, we’ll <strong>do a slightly different table calc configuration</strong>, but it will be the same again for both fields.</p><p>Right click the blue pill and select “Edit Table Calculation”. Select “Specific Dimensions”, make sure <strong>both your [Season] and [Team] are checked</strong>. <strong>Drag the [Team] field to the top</strong> and ensure <strong>“At the level” is also on [Team].</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/341/1*940ZRVuV5vTzJ2Lwa_xsoA@2x.png" /><figcaption>Configuring table calculation for map layers [Rows] and [Cols].</figcaption></figure><p>Again, feel free to add a custom sort in the “Sort order” control if you’d like. Just make sure to apply the same sort logic to both [Rows] and [Cols]. Once you’re done, you should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8ztv9z1Mgt6asgqWz42wGw@2x.png" /><figcaption>Tableau view after editing table calculations.</figcaption></figure><p>Take [Wins per Season Makepoint] and again drag it into the view to <strong>add another marks layer</strong>. This time, change your <strong>Mark Type to Area</strong>, drag <strong>[Team] to Detail</strong>, and <strong>[Season] to Detail</strong>.</p><p>Now that we have 2 map layers, we can go in the top toolbar and select <strong>Map → Background Maps → None</strong>. You should see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5HJHmz677cslfgDNpdigsA@2x.png" /><figcaption>Tableau view after removing background maps.</figcaption></figure><p>For the final calculation, I chose to add the team label in the top center of each team’s view.</p><pre>// Team Name Makepoint<br>// We&#39;re just plotting a single dot in space to write our label on<br>// I used 18 because it sits above the tallest wins<br>// I used 0.5 because we normalized our seasons between 0 and 1.<br>MAKEPOINT(18, 0.5)</pre><p>Add [Team Name Makepoint] as another map layer, change the <strong>Mark Type to Circle</strong>, under Color lower the <strong>opacity to 0%</strong>, add <strong>[Team] to Label</strong>, and change the <strong>text alignment to middle center</strong>. You should now see something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OSGNDpealtyeH2OPTK8cZw@2x.png" /><figcaption>Tableau view after adding team name labels.</figcaption></figure><p>The only thing left to do is clean-up! For layers like [Team Name Makepoint], make sure to click the dropdown and select <strong>“Disable Selection”</strong>. This prevents annoying hover effects from impacting your users.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/291/1*3IzKRrntnYhn_k0eXDSRUQ@2x.png" /><figcaption>How to disable selection on a map layer.</figcaption></figure><p>Then, touch up your axes, lines, headers, colors, and fonts as you wish. My final sheet looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dDqbJfVlyKbpsWbFwHZogg@2x.png" /><figcaption>Result after final clean-up steps.</figcaption></figure><h3>Wrapping It Up</h3><p>See, that wasn’t so bad! Of all the calculations that I think are worth bookmarking and referring back to, [Rows] and [Cols] are 100% my top 2. In fact, I’m going to write them in once more at the bottom of this blog for my own future reference:</p><pre>// Cols<br>(INDEX()-1)%[pCols]<br><br>// Rows<br>INT((INDEX()-1)/[pCols])</pre><p>I hope this tutorial is easy to follow and you learned a new trick to elevate your Tableau visualizations. See you next month!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4233b4dca912" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[FeizaData 2025: Year in Review]]></title>
            <link>https://medium.com/@feizadata/feizadata-2025-year-in-review-68e1736071d7?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/68e1736071d7</guid>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Fri, 02 Jan 2026 03:45:50 GMT</pubDate>
            <atom:updated>2026-01-02T22:31:17.461Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C2CkeLymlwXzBF2SjpLQuw.png" /></figure><p>As I reflect on 2025, it has been a pretty busy year. A lot changed, there were many wins along the way, and the year truly flew by. Since time isn’t going to slow down anytime soon, I think it’s important to pause and reflect every once in a while, and the new year feels like the best time to do that.</p><p>This post will not look like most of my blog content. Instead, I want to take a moment to reflect on a few things from the past year: my favorite vizzes I created in 2025, some of the vizzes from other authors that inspired me, a few of my biggest personal and professional wins, and finally a look ahead at my goals for 2026.</p><h3>10 Vizzes I’m Most Proud Of</h3><h4>Idle Visual Album Poster (January 10, 2025)</h4><p>In 2024, my best friend achieved a huge milestone by releasing his first album on streaming platforms. To commemorate this, I created a data-backed poster visualizing the notes played across each song. This viz was a fun passion project to work on and ended up being one of my favorite pieces I published in 2025.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9fhwM7tRkGZq_4qcTpOPwA.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/idlebyyupyiyvisualalbumposterfeizadata/IdleAlbumPoster">https://public.tableau.com/app/profile/blake.feiza/viz/idlebyyupyiyvisualalbumposterfeizadata/IdleAlbumPoster</a></figcaption></figure><h4>South Park #DataPlusMovies (January 20, 2025)</h4><p>As part of the #DataPlusMovies community project between IMDb and Tableau, I wanted to create a South Park-themed viz to participate and explore a show I genuinely enjoy. I spent way too much time in Figma designing the background image (probably longer than the analysis), but the end result made it worth it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p7BTsIVYvByKDD6fGAszuw.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/ComeonDowntoSouthParkDataPlusTV/DATAPLUSTV">https://public.tableau.com/app/profile/blake.feiza/viz/ComeonDowntoSouthParkDataPlusTV/DATAPLUSTV</a></figcaption></figure><h4>Impact of Industrialization by Income Group (March 10, 2025)</h4><p>While this viz isn’t particularly flashy, it’s built on a connected scatterplot that helped surface an interesting insight. By using size along each line to represent time, the x-axis is freed up to focus on the relationship between urban and rural populations across income groups. It‘s a fairly simple approach, but one that’s seldom used.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*23ZtD83ujWPcXd4-VgCdlw.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/ImpactofIndustrializationbyIncomeGroupMakeoverMonday/ImpactofIndustrializationbyIncomeGroup">https://public.tableau.com/app/profile/blake.feiza/viz/ImpactofIndustrializationbyIncomeGroupMakeoverMonday/ImpactofIndustrializationbyIncomeGroup</a></figcaption></figure><h4>Work Productivity Benchmarks (March 25, 2025)</h4><p>Map layers were clearly on my mind in March. I spent way too long trying to architect these striped bars (see my <a href="https://medium.com/retail-tableau-user-group/tableau-map-layers-magic-striped-bar-chart-67a4b7f544ea?source=your_stories_outbox---writer_outbox_published-----------------------------------------">tutorial blog post here</a> if you’re curious), but the battery-style effect ended up working really well for this productivity dashboard.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5TY4oJk-CetVYEOGxpFR4Q.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/ImpactofIndustrializationbyIncomeGroupMakeoverMonday/ImpactofIndustrializationbyIncomeGroup">https://public.tableau.com/app/profile/blake.feiza/viz/ImpactofIndustrializationbyIncomeGroupMakeoverMonday/ImpactofIndustrializationbyIncomeGroup</a></figcaption></figure><h4>Top Migration Destinations (April 6, 2025)</h4><p>Map layers again? Yes, but this time the technique <em>actually</em> made things easier. I used map layers to build the radar charts, combined them with a trellis layout to create the country grid, and added custom icons for each dimension to better communicate what each section represents. I had a lot of fun with this one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aoVK3KqSoE3L5NZAl6J9fA.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/TopMigrationDestinationsin2025MakeoverMonday/MM2025W14">https://public.tableau.com/app/profile/blake.feiza/viz/TopMigrationDestinationsin2025MakeoverMonday/MM2025W14</a></figcaption></figure><h4>#GamesNightViz Pokédex (May 9, 2025)</h4><p>I created this viz as my submission for the #GamesNightViz “Gotta Viz ’Em All” challenge. I knew I wanted to build something creative for this project, and Tableau animations had been top of mind at the time. By using map layers and manually tracing artwork for 20 different Pokémon, I was able to create a fun animated effect that really brought the dashboard to life. Definitely one of the more enjoyable builds from the year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FOUn_hSaXI0sMrw0RlKqzg.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/PokdexTableauEditionGamesNightViz/Pokdex">https://public.tableau.com/app/profile/blake.feiza/viz/PokdexTableauEditionGamesNightViz/Pokdex</a></figcaption></figure><h4>Cycles of Joblessness (August 20, 2025)</h4><p>When this Makeover Monday dataset was released, I wasn’t particularly excited. I didn’t know much about UK history, and at first glance the unemployment trend just looked like a random line chart. With some help from ChatGPT, I dug into the historical context and found anecdotes that helped explain why unemployment was rising or falling at different points in time. Combined with some Figma magic, the final result came together better than I expected.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bSSKQ_oTgerLH_NnfuflCg.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/CyclesofJoblessnessMakeoverMonday/MakeoverMonday">https://public.tableau.com/app/profile/blake.feiza/viz/CyclesofJoblessnessMakeoverMonday/MakeoverMonday</a></figcaption></figure><h4>365 Days of Play (August 27, 2025)</h4><p>So… map layers, again! This viz was inspired by a TikTok, and I took it as a challenge to recreate it in Tableau. I also wrote a blog post on this one, <a href="https://medium.com/dataai/behind-the-build-sports-radial-calendar-0ba37ee94910?source=your_stories_outbox---writer_outbox_published-----------------------------------------">check it out here</a> if you’re curious. Overall, this was one of the hardest things I built all year, but easily one of the most rewarding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-iU2MHh8PpEEsLjAv10nVg.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/365DaysofPlay/RadialSportsCalendar">https://public.tableau.com/app/profile/blake.feiza/viz/365DaysofPlay/RadialSportsCalendar</a></figcaption></figure><h4>Drug Harm in the UK (September 12, 2025)</h4><p>Once again, this chart type was not found on the “Show Me” shelf… Playing on the idea of getting “high,” I used a sigmoid function to create a droopy balloon-style chart for some of the most common drug types in the Makeover Monday dataset. This one was a fun experiment that panned out pretty well!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JKZbwNgVMbWJSfJjG4PWmA.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/DrugHarmintheUKMakeoverMonday/Feiza">https://public.tableau.com/app/profile/blake.feiza/viz/DrugHarmintheUKMakeoverMonday/Feiza</a></figcaption></figure><h4>F1’s Fiercest Rivalry (September 17, 2025)</h4><p>For a long-form viz, this one came together fairly quickly. As part of a case study, I had the opportunity to showcase my Tableau skills through a data project. I built out a narrative using a handful of chart types to tell the story of a rivalry between two F1 drivers, and it ultimately played a role in helping me land a new job.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/804/1*gS7Dskc9fZM3XJPnAf8WJw.png" /><figcaption><a href="https://public.tableau.com/app/profile/blake.feiza/viz/Formula1sGreatestRivalry/F1">https://public.tableau.com/app/profile/blake.feiza/viz/Formula1sGreatestRivalry/F1</a></figcaption></figure><h3>10 Vizzes That Inspired Me</h3><p>There’s no community that shares work quite like the Tableau community. I wanted to take a moment to highlight a handful of vizzes from 2025 that inspired me throughout the year, whether through their design choices, storytelling, or technical execution.</p><h4>Kings of Confection by Brian Moore (January 24, 2025)</h4><p>This viz makes me salivate a bit. Brian Moore does a fantastic job telling the story of candy in America through a long-form viz, and the design completely blew me away. I really enjoyed reading through this dashboard and learned a few things along the way too!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MBonhfYj3YxagHwd9tRlRA.png" /><figcaption><a href="https://public.tableau.com/app/profile/brian.moore7221/viz/KingsofConfection/KingsofConfection">https://public.tableau.com/app/profile/brian.moore7221/viz/KingsofConfection/KingsofConfection</a></figcaption></figure><h4>The Lions Prey by Chris Westlake (June 16, 2025)</h4><p>When Chris Westlake publishes a new viz, you already know it’s going to be worth your time. This rugby dashboard caught my eye immediately and does an excellent job encoding a ton of information into a single shape. The spirals are especially fascinating to me, and I’d love to use this as inspiration for a future viz.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hOgJHwRI7sENEAGjMD57MQ.png" /><figcaption><a href="https://public.tableau.com/app/profile/westlake.cjw/viz/TheLionsPrey_17500907162260/TheLionsPrey">https://public.tableau.com/app/profile/westlake.cjw/viz/TheLionsPrey_17500907162260/TheLionsPrey</a></figcaption></figure><h4>Super Mario Kart by Matt Huff (June 28, 2025)</h4><p>I really enjoyed Matt’s approach to this #GamesNightViz Mario project. His use of map layers to build these charts is clever and works really well in this dashboard. Combined with a strong themed design, this dashboard stands out as one of my favorites from the year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SS8h3KhbYuh6b0qw1SozWg.png" /><figcaption><a href="https://public.tableau.com/app/profile/matt.huff/viz/SuperMarioKart/Dashboard1">https://public.tableau.com/app/profile/matt.huff/viz/SuperMarioKart/Dashboard1</a></figcaption></figure><h4>Corruption Perceptions Index by Zyad Wael (August 7, 2025)</h4><p>Week over week, Zyad is one of the creators I most look forward to seeing participate in Makeover Monday. I liked this dashboard so much that I actually tried to recreate the design the following week with a similar dataset. A great example of a clean, effective dashboard.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mt_OFY9ukBvbXooiocAoDw.png" /><figcaption><a href="https://public.tableau.com/app/profile/zyad.wael/viz/CorruptionPerceptionsIndex-MakeoverMondayWeek32/Dashboard">https://public.tableau.com/app/profile/zyad.wael/viz/CorruptionPerceptionsIndex-MakeoverMondayWeek32/Dashboard</a></figcaption></figure><h4>Satisfying the appetite for sustainable seafood by Brittany Roseneau (August 23, 2025)</h4><p>Brittany’s long-form viz is one of my favorite examples of storytelling from 2025. I love the natural progression of the narrative, and the design and chart choices complement the story perfectly from start to finish.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6pFj38A_qHVSgpSJf97eUA.png" /><figcaption><a href="https://public.tableau.com/app/profile/brrosenau/viz/Satisfyingtheappetiteforsustainableseafood/Satisfyingtheappetiteforsustainableseafood">https://public.tableau.com/app/profile/brrosenau/viz/Satisfyingtheappetiteforsustainableseafood/Satisfyingtheappetiteforsustainableseafood</a></figcaption></figure><h4>Stock Performance by Seun Adeyemo (September 28, 2025)</h4><p>I’m still impressed that Seun took the time to learn the algorithm Tableau uses to render its default tree chart. With that knowledge, he was able to build a fully custom tree diagram with rounded corners, custom labels, and beautiful animations when filters are toggled. So cool!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-NE85etqhEFdkZmkvCxQEA.png" /><figcaption><a href="https://public.tableau.com/app/profile/bigxpt/viz/STOCKPERFOMANCE/unique">https://public.tableau.com/app/profile/bigxpt/viz/STOCKPERFOMANCE/unique</a></figcaption></figure><h4>Pawnee Parks and Recreation by Lisa Trescott (October 4, 2025)</h4><p>I’ve never seen a Tableau dashboard that feels so much like a real website. Once again, Lisa completely blows me away with her Tableau skills and delivers an incredibly well-designed exploratory dashboard that feels so unique.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zNpEKE2tsX7gmoo3a_Zlaw.png" /><figcaption><a href="https://public.tableau.com/app/profile/lisa.trescott/viz/PawneeParksandRecreation/Home">https://public.tableau.com/app/profile/lisa.trescott/viz/PawneeParksandRecreation/Home</a></figcaption></figure><h4>Mastering Tableau Series (#17) by Rubén Martínez (November 10, 2025)</h4><p>Rubén has over 20 dashboards in his “Mastering Tableau” series, and each one is chef’s kiss. They aren’t only beautifully designed, but they also include clear explanations for advanced techniques, chart types, and core fundamentals. Every time he posts a new one, I’m excited to see what he’s sharing next!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0Sn5BERrStWx2RrarugWaQ.png" /><figcaption><a href="https://public.tableau.com/app/profile/rub.nm/viz/MasteringTableau17_Tablecalculations2-Windowcomparisons/Dashboard">https://public.tableau.com/app/profile/rub.nm/viz/MasteringTableau17_Tablecalculations2-Windowcomparisons/Dashboard</a></figcaption></figure><h4>Playing with Food by Louis Yu (November 18, 2025)</h4><p>No matter the topic, Louis always finds a way to tie the Iron Viz theme back to video games. This dashboard is another great example of that! I really enjoyed the interactivity and design choices throughout this dashboard and found myself clicking around all the knobs and buttons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Se7cFh2PpfBDzeerGnpFng.png" /><figcaption><a href="https://public.tableau.com/app/profile/louisyu/viz/PlayingwithFoodIronViz/Evolution">https://public.tableau.com/app/profile/louisyu/viz/PlayingwithFoodIronViz/Evolution</a></figcaption></figure><h4>Nutrition, Energy Sources, and Ingredients of Dim Sum by Kevin Wee (November 21, 2025)</h4><p>When Iron Viz season came around, Kevin was one of the artists I was most excited to see. After last year’s Broadway-themed qualifier submission, I was really curious to see what he would come up with this year. The design and chart choices in this dashboard are a true masterclass!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0lQkOnaBW4U4IWNB79HyLA.png" /><figcaption><a href="https://public.tableau.com/app/profile/kevin.wee/viz/NutritionValuesofDimSumIronVizedition/Dashboard">https://public.tableau.com/app/profile/kevin.wee/viz/NutritionValuesofDimSumIronVizedition/Dashboard</a></figcaption></figure><h3>Reflection: Wins in 2025</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uhx4YgHHnXD1RQuA" /><figcaption>Photo by <a href="https://unsplash.com/@sagarp7?utm_source=medium&amp;utm_medium=referral">Sagar Patil</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I had some big wins and checked off some goals this year. One of those goals was to publish a blog post every month in 2025. If you are reading this, I’m so thankful for your time. It makes the whole writing experience worth it.</p><p>As I reflected on the year, a few other themes stood out:</p><h4>TC25</h4><p>I attended my second Tableau Conference this year in San Diego and had the opportunity to speak for the first time alongside Matt Huff. It was an incredible experience, and seeing a standing-room turnout was something I will never forget. We also led an in-person meetup for the Retail and Consumer Goods Tableau User Group with Ojoswi Basu. It was a great experience all around, and I am already looking forward to building on this momentum at TC26.</p><h4>Community Involvement</h4><p>2025 was the first full year of the #RetailTUG. We hosted 6 events this year with some incredible guest speakers, including one hybrid session. The user group continues to trend upward, and I am excited to keep delivering strong, relevant content in the year ahead.</p><p>Toward the end of the year, I also announced that I will be joining the #MakeoverMonday team. As a frequent participant in this project over the past few years, I am incredibly grateful for the opportunity and excited to help curate datasets each week that could use a “makeover”.</p><p>Lastly, I am proud to share that I was named both a 2025 DataFam Rising Star and a 2025 Tableau Ambassador. Both of these are community-nominated recognitions, and it means a lot to be acknowledged by a group I have looked up to for years. I plan to keep giving back in 2026!</p><h4>Professional Development</h4><p>I am incredibly happy with the progress I’ve made with Tableau, and a large part of that growth is thanks to Next-Level Tableau. I have been training with Andy Kriebel for the past two years and am excited to continue that journey in 2026. His mentorship has been invaluable as I’ve grown technically and expanded my network.</p><p>This year also marked a major career transition for me, as I moved from HEB to a role at Apple as a Senior Data Visualization Engineer. It has been an exciting next step, and I’m grateful for the experiences that led me here.</p><h3>Goals Going Into 2026</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VzixlWZeVQ2n0lUi" /><figcaption>Photo by <a href="https://unsplash.com/@glenncarstenspeters?utm_source=medium&amp;utm_medium=referral">Glenn Carstens-Peters</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Even though there is some well-documented psychological research suggesting that sharing your goals can actually make you less likely to achieve them (the short version being that our brains release a sense of “success” just by saying them out loud), there are still a few things I want to call out as I look ahead to 2026.</p><h4>Continuing to Blog Monthly on FeizaData</h4><p>I’ve found that publishing one blog post a month is a pretty healthy cadence. It gives me enough time to be thoughtful about each post while still staying consistent. I still have plenty of ideas I want to explore here and have not run out of gas yet. :)</p><h4>Progressing as a Community Leader</h4><p>Through blogging, creating vizzes, speaking at TC26, and continuing to lead community user groups and projects, I hope to keep growing as a community leader in the Tableau and data visualization space. More than anything, I want to stay engaged, keep learning, and help create spaces where others feel encouraged to share their work as well.</p><h4>Trying Something New?</h4><p>I have a tendency to sign myself up for more than I can chew (I even signed up to run a 50K in May…) At some point in 2026, I would like to try my hand at livestreaming some viz sessions. Makeover Monday feels like a natural place to experiment with this, using a timeboxed format to build something each week. It would be similar in spirit to Andy Kriebel’s old “Watch Me Viz” series on YouTube. We’ll see if I actually commit to the idea now that it’s in writing…</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*JWZwQybnmHVfTUF9.jpg" /><figcaption><a href="https://www.youtube.com/watch?v=AB61I9VK7Xs">Into The Wild | Happiness Only Real When Shared — YouTube</a></figcaption></figure><p>To wrap up the year and this post, I just want to say thank you if you’re reading this. As much as it makes me happy to write posts about topics that excite me, it means the world to know there are people on the other end that are consuming it. I hope you stick with me in 2026!</p><p>-Blake Feiza</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=68e1736071d7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Easiest DZV Tutorial of All Time]]></title>
            <link>https://medium.com/@feizadata/the-easiest-dzv-tutorial-of-all-time-a81359dcb088?source=rss-5da05b3c7452------2</link>
            <guid isPermaLink="false">https://medium.com/p/a81359dcb088</guid>
            <category><![CDATA[datafam]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[dynamic-zone-visibility]]></category>
            <category><![CDATA[tableau]]></category>
            <dc:creator><![CDATA[Feiza Data]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 03:17:41 GMT</pubDate>
            <atom:updated>2025-12-02T03:17:41.183Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fbLQ8xdeWcue9Cdo0UaGUA.png" /><figcaption>The Easiest DZV Tutorial of All Time</figcaption></figure><p>I remember early in my Tableau journey hearing about Dynamic Zone Visibility (DZV) and thinking it must be some pretty advanced stuff.</p><p>I’m writing to prove to you it’s actually one of the easiest interactivity features in Tableau.</p><p>I’ll break the process into 3 easy steps and mention some common misconceptions along the way. If you’re already familiar with DZV, I’ll also add an advanced technique that might solve that “impossible” problem you’ve been stuck on.</p><h3>DZV in 3 (Simple Steps)</h3><h4>1. Create a Parameter</h4><p>In all DZV use cases, the first step is to create a parameter. The type doesn’t matter here, because step 2 will reference this parameter inside a calculation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*wDDPwg3USJysJjEaGz67WA.gif" /><figcaption>Step 1. Create a Parameter</figcaption></figure><h4>2. Create a Calculated Field</h4><p>A parameter behaves like a light switch with no wiring by itself. It doesn’t control anything until we pair it with a calculation. For DZV to work, we need a calculated field that evaluates to exactly TRUE or FALSE across the entire workbook.</p><p>The key idea is that parameters always return a single value globally. That makes them perfect for DZV. Filters do not work here because they can vary row by row or sheet by sheet.</p><p>In this example, I create a simple calculated field like this:</p><pre>// DZV. Show Chart<br>[DZV Parameter] = &#39;Show&#39;</pre><p>Observe that this calculation will always evaluate to either TRUE or FALSE. Also, because parameters always have one global value, this calculation’s result will always be ONLY TRUE or ONLY FALSE at any time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*OwyoOQV7xhnYRvvcuYBOKg.gif" /><figcaption>Step 2. Create a Calculated Field</figcaption></figure><blockquote>Note that in this example, my [DZV Parameter] has 3 values; “Show”, “Hide”, and “Etc”. The calculation I’m using will show the chart when the parameter is “Show” and hide for all other values. We can modify the logic here in a couple ways depending on the goal:</blockquote><pre>// Only Show the Chart on &quot;Show&quot;<br>[DZV Parameter] = &quot;Show&quot;<br><br>// Only Hide the Chart on &quot;Hide&quot;<br>[DZV Parameter] != &quot;Hide&quot;<br><br>// Show the chart for a specific set of options<br>[DZV Parameter] IN (&quot;Show&quot;, &quot;Etc&quot;)</pre><h4>3. Edit on Dashboard</h4><p>Now that our calculation from Step 2 exists, we don’t even need to add it anywhere. Tableau simply stores it in the workbook and recognizes it as a “DZV-eligible” field. We can head to our dashboard and reference it directly.</p><p>DZV is a property you can apply to any dashboard object. To modify this property, select an object, then go to <strong>Layout &gt; Control visibility using value</strong>.</p><p>If your calculated field was set up correctly, it will appear in the DZV drop-down. If it does not appear, revisit the calculation; this is usually because the logic doesn’t resolve to either TRUE or FALSE across the entire workbook.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*IARsrxPzJ4x0KtT--rrs5A.gif" /><figcaption>Step 3. Edit on Dashboard</figcaption></figure><p>If everything works correctly, you’re done! You can add as many calculations as you need to cover different user experience scenarios, and you can apply the same calculation to any number of objects or dashboards.</p><p>One final note: if you are using a Boolean parameter, you can actually connect DZV directly to the parameter without creating a calculated field. This only applies to true Boolean parameter types, but it is a helpful technical shortcut to know.</p><h3>Bonus Technique: When “Easy DZV” Won’t Cut It</h3><p>Sometimes, it can be surprisingly difficult to create a calculation that is globally TRUE or FALSE and still meet the requirements for your DZV use case. To show what I mean, here is an example from Superstore:</p><p>Suppose I want to look at the number of yearly orders by ship mode for a selected manufacturer. The dashboard below shows the result for AT&amp;T, and everything looks fine because there are orders in every ship mode.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B4NmrduhVEC6NnVvsXPCJQ@2x.png" /><figcaption>Bonus Technique with all 4 sheets available.</figcaption></figure><p>When I switch my [Manufacturer Parameter] to Ativa, there are no First Class orders. This leaves an empty space in the layout where that bar would have been, and I would prefer that First Class disappear entirely so the remaining ship modes have more room.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y4GBZNLfLiUd1wNaYOywbw@2x.png" /><figcaption>Bonus Technique with sheet missing data.</figcaption></figure><p>This is where DZV becomes a bit trickier.</p><p>To control visibility for each ship mode independently, we need calculated fields (one for each ship mode) that always resolve to TRUE or FALSE across the entire workbook. These calculations need to understand whether the selected manufacturer has at least one qualifying record.</p><p>The key is using fixed LODs.</p><p>I created the following calculations (one for each of my four ship modes):</p><pre>// DZV. First Class Orders?<br>{ COUNTD(<br>    IF [Manufacturer] = [Manufacturer Parameter] AND [Ship Mode] = &#39;First Class&#39;<br>    THEN [Order ID] <br>    END<br>  ) &gt; 0 <br>}<br><br>// DZV. Same Day Orders?<br>{ COUNTD(<br>    IF [Manufacturer] = [Manufacturer Parameter] AND [Ship Mode] = &#39;Same Day&#39;<br>    THEN [Order ID] <br>    END<br>  ) &gt; 0 <br>}<br><br>// DZV. Second Class Orders?<br>{ COUNTD(<br>    IF [Manufacturer] = [Manufacturer Parameter] AND [Ship Mode] = &#39;Second Class&#39;<br>    THEN [Order ID] <br>    END<br>  ) &gt; 0 <br>}<br><br>// DZV. Standard Class Orders?<br>{ COUNTD(<br>    IF [Manufacturer] = [Manufacturer Parameter] AND [Ship Mode] = &#39;Standard Class&#39;<br>    THEN [Order ID] <br>    END<br>  ) &gt; 0 <br>}</pre><p>Each ship mode chart gets its own calculation that checks whether there is at least one [Order ID] matching both the manufacturer and the ship mode requirements. If there is at least one match, the calculation returns TRUE. If not, it returns FALSE.</p><p>By wrapping the logic in a FIXED LOD, the outcome becomes global and stable, which is exactly what DZV requires.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*qAp785l_lHxl_pDd3auz5w.gif" /><figcaption>Adding Bonus Technique DZV Calcs for solution.</figcaption></figure><p>Variations of this pattern have solved several otherwise challenging layout problems for me, and I hope it helps you with yours.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a81359dcb088" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>