<?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 Bjorn Austraat on Medium]]></title>
        <description><![CDATA[Stories by Bjorn Austraat on Medium]]></description>
        <link>https://medium.com/@bjorn.austraat?source=rss-6dd9a05cc909------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Y5mntCTHxW_dxsVX73RnIA.jpeg</url>
            <title>Stories by Bjorn Austraat on Medium</title>
            <link>https://medium.com/@bjorn.austraat?source=rss-6dd9a05cc909------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 04:12:40 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@bjorn.austraat/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[Movie Revenue Prediction — What can data tell us about the “magic formula” for box office success?]]></title>
            <link>https://medium.com/@bjorn.austraat/movie-revenue-prediction-what-can-data-tell-us-about-the-magic-formula-for-box-office-success-ab3fc2cbcd7a?source=rss-6dd9a05cc909------2</link>
            <guid isPermaLink="false">https://medium.com/p/ab3fc2cbcd7a</guid>
            <category><![CDATA[data-science-courses]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Bjorn Austraat]]></dc:creator>
            <pubDate>Sun, 06 Dec 2020 15:52:40 GMT</pubDate>
            <atom:updated>2023-05-31T19:26:04.359Z</atom:updated>
            <content:encoded><![CDATA[<h3>Movie Revenue Prediction — What can data tell us about the “magic formula” for box office success?</h3><p>Can we build a machine-learning model to predict a movie’s revenue potential? I used the <a href="https://www.kaggle.com/tmdb/tmdb-movie-metadata">Kaggle TMDB 5000</a> data set to find out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DywgVk0WQX66Io_boucHwA.jpeg" /></figure><p><em>[This article was submitted as part of the Capstone Project for a </em><a href="https://www.udacity.com/course/data-scientist-nanodegree--nd025"><em>Udacity Data Science Nanodegree</em></a><em> program December 2020 ]</em></p><h4>The challenge</h4><p>Since the stakes are often very high, studio bosses, movie producers, and many others in the film industry have been looking for a surefire way to turn their investment into big box office returns since the earliest days of the industry.</p><p>A movie’s box office and critical success, like that of any creative endeavor, is a combination of a large number of tangible and intangible elements. The onscreen chemistry of the main characters, the director’s vision and execution, production values / special effects, and the movie’s release timing are just a few of the components that can turn low- or no-budget indy flicks into huge blockbusters, or doom an expensively produced studio vehicle or even a ‘surefire’ sequel to crater at the box office.</p><h4>Is there a “formula for movie success?”</h4><p>Can past movie release data provide us with a ‘formula’ to increase our chance of commercial success? Are there ‘winning combinations’ of actors, topics, directors, and genres? This project explores a <a href="https://www.kaggle.com/tmdb/tmdb-movie-metadata">database of 4,800 movies </a>to first create a visual and quantitative intuition for potential patterns present in movie data and then attempts to create a machine-learning model to predict a movie’s revenue.</p><p>To measure initial model performance and incremental improvements from subsequent tuning passes I used the <a href="https://scikit-learn.org/stable/modules/model_evaluation.html#r2-score">R2 score</a> as well as the <a href="https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_squared_log_error.html">Root Mean Square Log Error</a>.</p><h3>Data load, initial exploration, and feature engineering</h3><p>Upon initial load of the dataset, we see a pretty diverse set of features, including complex types like dictionaries (for cast/crew/genres, etc.).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ykx63VIOgpathgm2WHLcBA.jpeg" /></figure><p>I created a handful of custom functions to convert dictionaries into lists of strings and custom objects with bespoke properties to make later processing easier. For example, the below code creates a custom ‘cast_member’ class and then converts the given cast member dictionary into an iterable list of custom cast-member objects.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uWSjk8EqQDADAoOS5Avg5Q.jpeg" /></figure><p>The class and function are then used in a vectorized dataframe operation to populate the appropriate column with my iterable custom class list (vs. the given JSON dictionary).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pk_vQnimYR2HNVo5IjBabg.jpeg" /></figure><p>Next, I used a mix of lambda and custom functions to generate additional polynomial features.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SKEVCZ14wG_GU6AxmiVKtg.jpeg" /></figure><h4>Trouble ahead…serious ‘plot holes’ in my data</h4><p>Upon closer inspection of several key elements, I realized that the data set had quite a few gaps. Of top concern were missing or implausibly low numbers for revenue and budget, especially since ‘revenue’ was going to be the target variable that I needed to train my machine learning model to predict accurately. Specifically, there were 1,069 movies with an implausibly low budget and 1,451 movies with implausibly low revenue.</p><p>There are a number of ways to deal with missing data in machine learning. I could have tried to fill the missing budget values with the mean of all budgets, or create a mini-machine-learning model to try to predict the budget from other dimensions, or — worst choice—tried to manually populate the missing 2,500 values.</p><h4><strong>BeautifulSoup to the rescue!</strong></h4><p>I decided to create a little ‘screen-scraping’ utility with <a href="https://www.crummy.com/software/BeautifulSoup/bs4/doc/">BeautifulSoup</a>, a Python library for pulling data out of HTML and XML files. The utility navigates to the appropriate IMDB.com page for the movie in question and pulls budget and revenue data if available. Unfortunately, the source database did not contain usable IMDB movie IDs that could be used to land directly on the appropriate page. In the absence of unique, URL-mappable movie IDs, I performed a URL-encoded search function first, then selected the most likely result link candidate, followed that link, and then parsed the resulting page for budget/revenue information.</p><p>Step 1: Find the right page on IMDB.com using a search for movie title and year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G-svKVTfpY_gXjnBrG8IqA.jpeg" /></figure><p>Step 2: Parse through the resulting page and extract useful bits.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*axrZzOGz-DyKpJwDk7x2ZA.jpeg" /></figure><p>Letting BeautifulSoup perform thousands of programmatic ‘click, cut, and paste’ operations turned out to be a huge timesaver.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/833/1*c0dTa6scULDk3KbWLN4w6g.gif" /></figure><p>The results were satisfying as well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-WiD09fOdHgvsGQps-ALWA.png" /></figure><p>This web scraping exercise cut the number of defective entries between 66% and 78%!</p><h3>Data visualization</h3><p>Now it was time to visually explore both the quantitative and categorical aspects of the data set.</p><h4>Categorical data visualization</h4><p>To start, I wanted to use the WordCloud library to generate word cloud images of genres and keywords to get a better intuition of the dominant expressions used in both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*sjKip8RaOnKExAKGD54kjA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*2rLT3Py9VxfLhEpa8AmH8A.png" /></figure><p>As an additional deep-dive, I also generated word clouds of ‘keywords by genre.’</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*Cgs_lIr8nL0fdMwYGN-lMQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*9EVD2PeiSmWNDhu-Kqankg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*kmQjZlZUOlfAWFyufVKCUg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*whnoAtUZ6ZPNe4ByklkZ5Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*0oqLyfbKh8RcvASoXSiFGg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/572/1*-rD9z_7nWNFUAB0lmldc5w.png" /></figure><p>To get a feel for the quantitative distribution of both genres and keywords, I generated their respective histograms for all genres and the top 100 keywords.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cdwRc3VeLwEjUdgMGNDpgg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2XFz7FV3f4NZEDDkpwJ-3Q.jpeg" /></figure><h4>Quantitative data visualization</h4><p>To start the quantitative exploration, I first wanted to get a good sense of budgets vs. revenues over time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GsdpZfUHCk3sFr4Sy0jfqw.jpeg" /></figure><p>The steep drop-off to the right is most likely due to an incomplete data set for the last observed year. Also interesting is the spikiness of the data in the 1940s. Could indicate data defects or a few smash hits?</p><p>I also wanted to find out how individual quantitative features correlated. This could be of interest in a later feature tuning pass.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/784/1*L3jETsOLUGTYUMKOYPP5-Q.png" /></figure><p>No big surprises there…a quick deep dive into correlations between “budget and revenue” as well as “budget and popularity” gave additional visual confirmation of the intuitive connection between these elements.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*QEdAxxfSrADfdvZ_J8l0Pw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*pLh6MFd_HmQrAWh4Umln6A.png" /></figure><p>For the Popularity vs. Revenue scatterplot, I adjusted the feature scale via an np.log1p() transformation to avoid an unnaturally flat trend line given the wildly different ranges in popularity and revenue dimensions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LLyAtfG09rsbsx-C4hRrQw.png" /></figure><p>Next, I wanted to get a better understanding of the geographic distribution of movie production. Unsurprisingly, the US emerges as the “800-pound gorilla” in the 70s.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/836/1*zl3D-fwPjmwCCO3tEQplFA.png" /></figure><p>Last, but not least, I investigated the ‘commercial leaders’…which directors and actors are the most prolific with commensurate commercial results?</p><p>The ‘usual suspects’ quickly emerged in the visualization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Je5R6nPjYoNHvCpGfMslIg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5051MN3nWWnZ3Bw6Ttp1dw.jpeg" /></figure><h3>Building the Machine Learning Model</h3><p>I chose a Light Gradient Boosting Machine Regressor for this project given its modest memory requirements and good overall performance.</p><h4>Additional feature engineering —Custom one-hot encoding by rank and revenue impact</h4><p>There are thousands of actors, directors, and keywords in the database. One-hot encoding all of them would create a very large, but exceedingly sparse feature matrix. Instead, I developed a custom one-hot encoder that is configurable with an additional rank-sorted data structure and can one-hot encode only the top-sorted <em>n</em> entries.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wMRxyqrIK2z8V1mGRerrSA.jpeg" /></figure><p>The custom one-hot encoder converted the more challenging dimensions (keywords, genres, cast, directors), leaving only elements that could be transformed using standard pipeline elements.</p><h4>Pipeline generation and initial fit</h4><p>I continued the setup by combining a one-hot transformer, a TFIDF Vectorizer, and another np.log1p transformer into a ColumnTransformer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gFeyBdliNE9wQxRgZgSTnw.png" /></figure><p>Next, I configured the hyperparameters for the LGBMRegressor and then put the whole pipeline together.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u51j6BRxA_EhLhdvFmEZGw.png" /></figure><p>Wrapping the model in a target transformer turned out to be critical. In early experiments, the R2 score was only about 0.006 since the huge range of the target variable from 5,000 to 2.7B made a proper fit almost impossible.</p><p>After a cross-validation pass to examine model portability, I then performed an initial training fit to get preliminary results.</p><p>The model performed quite well, with an R2 Score of 0.942925 and an RMSLE of 1.379354.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*TzpLoZm2aUlG4cUl0X9ciA.png" /></figure><h4>Model tuning</h4><p>In an attempt to further optimize results, I configured a Bayesian tuner for a subset of the pipeline’s hyperparameter space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bQmI8TFBkEs7UfCKduagOg.png" /></figure><p>After completion, the tuner suggested changes to the hyperparameters. Of particular interest were the changes likely to reduce overfitting including a reduction in the number of leaves, an increase in the number of estimators, an increase in the maximum depth, and the suggested change to the feature depth of the TFIDF vectorizer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CwrSg5CrgbUwU7nYpx7hUQ.png" /></figure><p>Using the resulting hyperparameters resulted in a slight improvement of the R2 score to 0.949329 and a slight decrease of the RMSLE to 1.301460.</p><h3>Conclusion and reflection</h3><p>The goal of this project was to determine whether historical movie data can be used to predict revenue on a test set or for future movies. I believe that given the high R2 score, there is a definitive role for machine learning as <strong>one of many</strong> inputs into the decision pipeline for studio executives.</p><p>Naturally, there are also many intangibles that contribute to the box office success of an individual release that cannot be captured adequately by pure data including on-screen chemistry of the lead actors, the competitive release schedule of other studios, external disruptions (e.g. the current Covid-19 pandemic), weather, etc.</p><h4>Challenges</h4><p><strong>Data quality: </strong>The source data set of approximately 5,000 movies was missing close to 30% of revenue figures and 20% of budget figures. As the dataset was also lacking an unambiguous URL-resolvable way of looking up the missing values, I created a ‘screen-scraping’ workaround using a URL-encoded search against the IMDB.com site (starting with the movie title and release year) and subsequently parsing the HTML results for budget and revenue data.</p><p><strong>Data format and feature variety: </strong>The original data set contained JSON encoded objects mixed in with strings, floats, and integers that required extensive pre-processing to create more convenient representations, especially of the JSON objects. In addition, there were natural-language features that required vectorization to become useable in a regression model (TFIDF Vectorizer was chosen in this notebook).</p><p><strong>Feature scale: </strong>The target feature of ‘revenue’ ranges from a few thousand dollars to 2.7B dollars. Similarly, the ‘budget’ feature ranges from several thousand to several hundred million dollars. This required rescaling of both features and target data points. In an initial attempt (before re-scaling the target variable), I achieved a measly R2 score of only 0.006 given the huge distribution of revenue.</p><p><strong>One-hot encoding of cast and crew based on revenue relevance: </strong>Traditional one-hot encoding would have a) resulted in a huge, extremely sparse matrix of thousands of actors and directors and b) not necessarily provided actors that are more strongly correlated to box-office success. I decided to create my own one-hot encoder that could take a ‘control dataframe’ as an additional input and only one-hot encode actors and directors high on the ‘commercial impact scale’, i.e. individuals with proven box-office success.</p><h3>Potential Improvements</h3><p>There are numerous aspects of this project that hold opportunities for additional growth and insights.</p><p><strong>Model alternatives, metric alternatives, and more extensive hyperparameter tuning: </strong>Additional model architectures could be evaluated to see if e.g. XGB results in superior performance. The selected model could also be more extensively tuned across a wider hyperparameter search space. In addition, we should consider other potential quality metrics of the regression outcomes like RMSE, RMSLE, etc.</p><p><strong>Additional data augmentation and changes to imputation mechanisms: </strong>Additional manual lookups or alternative automated screen-scraping attempts can be made to complete the remaining budget and revenue gaps in the data set. A machine learning approach could also be taken to impute any remaining gaps (vs. the simplistic ‘mean()’ approach chosen above).</p><p><strong>Additional feature engineering: </strong>Additional polynomial features could be generated to unearth other revenue-relevant aspects of the composition of cast &amp; crew, seasonality of release date, etc.</p><p><strong>Improved NLP processing: </strong>The current transformation using a single TFIDF vectorizer could be enhanced with additional steps including more sophisticated lemmatization, part-of-speech selection, and other steps.</p><p>If you would like to run your own experiments with different parameters, feel free to download the data, Jupyter notebook, and Python source code from <a href="https://github.com/NYCDS2020/DSND-Capstone-TMDB5000">Github here</a>.</p><p><a href="https://www.linkedin.com/in/bjornaustraat/">Bjorn Austraat</a> is SVP and Head of AI Acceleration at Truist and specializes in Agile AI at Enterprise Scale and Machine Learning and Deep Learning Innovation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ab3fc2cbcd7a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Zen of AI Innovation: How to bring order to your AI portfolio and build a solid AI roadmap in…]]></title>
            <link>https://medium.com/@bjorn.austraat/the-zen-of-ai-innovation-how-to-bring-order-to-your-ai-portfolio-and-build-a-solid-ai-roadmap-in-f34c45d03e09?source=rss-6dd9a05cc909------2</link>
            <guid isPermaLink="false">https://medium.com/p/f34c45d03e09</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ai-implementation]]></category>
            <category><![CDATA[enterprise-ai]]></category>
            <category><![CDATA[ai-innovation]]></category>
            <category><![CDATA[4d]]></category>
            <dc:creator><![CDATA[Bjorn Austraat]]></dc:creator>
            <pubDate>Wed, 21 Oct 2020 12:51:38 GMT</pubDate>
            <atom:updated>2023-05-31T19:27:02.316Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Zen of AI Innovation: How to bring order to your AI portfolio and build a solid AI roadmap in 3 easy steps</h3><p>Innovate faster, more productively, and with greater confidence</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c4yWFW4KYHYghXcXc41YVA.jpeg" /></figure><p>AI is moving from experimentation to industrialization at a staggering speed. Many organizations are struggling with ‘cat herding’ disparate AI innovation initiatives into a cohesive whole and balancing short-term payoffs with long-term transformation.</p><p>Here are some useful tips &amp; tricks to help you <strong><em>generate high-quality ideas</em></strong>, logically <strong><em>organize them into a well-balanced AI portfolio</em></strong>, and ultimately <strong><em>generate a coherent enterprise roadmap</em></strong>.</p><h3>Step 1: Generate high-quality candidate ideas</h3><p>Filling your AI innovation funnel with high-quality ideas is easier said than done. What’s a science experiment and what’s potentially the ‘next big thing?’ To get a holistic picture early on in the innovation lifecycle, I recommend the 4D Framework for vetting candidate ideas. The 4 Ds are <strong>Data, Dollars, Deployability, and Differentiation</strong></p><p><strong>Data — Lifeblood of all AI — </strong>Do you have data sets available, either as raw or already labeled or feature-engineered sets that can be used to train machine learning models? Is the data either readily understandable by data scientists or do you have subject matter experts that can provide this clarity for — often cryptically named — tables and columns? What are the regulatory constraints on data usage and will physical data gravity potentially become an issue for very large data sets?</p><p><strong>Dollars — Business Benefits Matter —</strong>How will your AI solution produce benefits? Is it focused on operational expenses and bottom-line savings or are you targeting top-line growth? Can you tell a simple value chain story — for instance, an AI-powered screen pop system in the contact center will reduce average handle time by 2 minutes on 10,000 calls per day for a savings of $20,000 (assuming a $1/minute cost)? How will you instrument, measure, and dashboard the system to prove AI impact? If your value case depends on generating additional share of wallet or market, how will you design your A/B experiment to isolate AI impact and prove business benefits independently of many other confounding factors?</p><p><strong>Deployability — Avoiding ‘stranded models’ — </strong>Think of your fantastic ML model like a high-performance race car engine. Certainly a feat of modern engineering, but utterly useless without the other bits that are required to win a race — including wheels, gas, a windshield, struts, brakes, a pit crew, etc. Early on, consider the downstream dependencies and other enterprise systems that have to be in place and ready for model outputs to produce benefits. Beyond technical integration, also ensure schedule alignment as upgrade or maintenance cycles for large enterprise systems are not always flexible enough to incorporate AI-driven changes.</p><p><strong>Differentiation — To AI or not to AI— </strong>Avoid blind spots and AI-biased thinking by asking yourself early on if AI is indeed the right tool for the job. Maybe a different approach produces value more quickly? Consider alternatives like Robotic Process Automation (RPA), or even spreadsheet-based or rules-based solutions. Maybe there is an existing third-party vendor or legacy model that can get you 80% there. Sometimes it’s not a technology question at all but a matter of team building and organizational change to address the root cause of an issue rather than fixing it with high-tech after it’s already occurred.</p><h3>Step 2: Get clarity about your innovation type</h3><p>Great! Your brainstorming sessions have generated a large number of high-quality innovation candidates using the 4D approach. Now what? To further prioritize your ideas and shape them into a well-structured portfolio, I recommend the “innovation 2x2” approach below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3TI9dUpl2xet_6ZbCZjAoA.png" /></figure><p>The dimensions are simple — capability and demand. Capability expresses your organization’s ability to execute and demand expresses the demand signal from internal and external customers. On both axes, values can range from ‘certain’ to ‘uncertain.’</p><p>We can now map innovative AI ideas into quadrants…from incremental to disruptive.</p><p>Let’s start with <strong>certain demand</strong> and <strong>certain capability</strong>. We know we can do it and we know our customers want it. This is where <strong>incremental </strong>innovation lives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U_OPHmql6kqLHXMlKGEdRA.png" /></figure><p>In AI terms, it’s typically a point solution that makes existing processes slightly better, like a more accurate OCR system based on neural networks, a better voice assistant, etc.</p><p>In the next quadrant, we look at <strong>certain demand</strong> encountering <strong>uncertain</strong> <strong>ability </strong>to deliver. It’s possible that AI capabilities required to meet customer demands are still emerging, or maybe they are and will remain outside of the strategic focus area for the enterprise or department. This type of innovation is best tackled with a competent <strong>partner</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NFw4qrQKX_eY-b3ofaofvA.png" /></figure><p>Here we rely on outside parties to deliver AI innovation and systems — something that is very common in AI. An example might be a third-party AI chatbot vendor that delivers full end-to-end capabilities for designing, building and running cross-channel bots.</p><p>When our <strong>ability</strong> to execute is <strong>certain</strong> but is paired with <strong>uncertain</strong> customer <strong>demand</strong>, we can experiment with low risk. If it works out, great, if it doesn’t, no big deal. This is where <strong>fast win/fail innovation</strong> projects live.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SFiqt6imTEDaMLhjZYC7MQ.png" /></figure><p>In AI, this might be a better marketing targeting model based on innovative algorithmic approaches. The model can be readily A/B tested, if it works it delivers a lift in revenue, if it doesn’t, it can be turned off after running a thorough experiment.</p><p>And lastly, the area that is most often associated with innovation — truly <strong>disruptive innovation</strong>. We are not sure we can do it, and we are not sure customers want it, but if the bet pans out, it will deeply transform a process, sector, or a whole industry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fS_-0q8rhAEmjAAohXr4cA.png" /></figure><p>Truly safe and reliable self-driving cars would fall into this category. Similarly, a form of robust general artificial intelligence would certainly be extremely disruptive and transformative.</p><h3>Step 3: Generate a sequenced roadmap</h3><p>Now that we have a well-sorted portfolio full of viable ideas, what should we do first? By overlaying time horizons, we can generate a logical project roadmap to give us a good sense of near-term, medium-term and long-term execution priorities.</p><blockquote>Note: Allocating ideas to Horizons 1, 2, and 3 as shown below works in many cases, but of course there will be exceptions where it’s perfectly reasonable to tackle a ‘partner’ or ‘disruptive’ project first. Consider the proposed sequence a rule of thumb with lots of wiggle room, not a law set in stone.</blockquote><p><strong>Horizon 1: </strong>Well-qualified 4D candidates that are incremental in nature can be prioritized into Horizon 1 initiatives. These typically deliver reliable and relatively short-term payoffs with business cases that are more readily explainable and accessible. An excellent way to ‘pay as you go’ in innovation and provide momentum for more aspirational and longer-term initiatives.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pDcEC2DmzVyIAkd3B-P77w.png" /></figure><p><strong>Horizon 2:</strong> Initiatives with a higher degree of uncertainty often fit into planning Horizon 2 as partner selection and vetting can be time-consuming. Fast win/fail experiments may in turn depend on having built a certain amount of credibility, especially if the experiment turns into a fast fail. This is why having a solid foundation of incremental wins in Horizon 1 and deferring these experiments Horizon 2 can be a more sustainable strategy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1vpXWsVxuljA0kBuyCWGzg.png" /></figure><p><strong>Horizon 3:</strong> Lastly, the most uncertain of projects — that often also rely on a top-down business model and aim to unlock whole new segments and markets, are best served with a longer planning horizon. This allows an enterprise to build up experience at executing AI projects…an organizational ‘muscle memory’ so to speak, combined with lots of credibility earned with earlier Horizon 1 &amp; 2 successes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cgw5bhChmxH80JZ-oSa3Eg.png" /></figure><p>I hope that you will find the concepts and frameworks in this post useful when launching AI ideas in your own organization. To start, the <strong>4Ds</strong> will boost cross-team collaboration, ideation speed, and productivity throughout the innovation journey. Next, the <strong>innovation portfolio</strong> view provides peace of mind by delivering a holistic perspective of diverse solution candidates and ensuring a balanced approach that fits your organization’s risk appetite. And together with the 4Ds and a solid portfolio view, structuring a long-term transformation journey into logical <strong>horizons</strong> allows you to achieve incremental goals without losing sight of the bigger picture.</p><p>Thank you for reading and I wish you great success, clarity, and ‘AI Zen’ on your own AI innovation journey!</p><p><a href="https://www.linkedin.com/in/bjornaustraat/">Bjorn Austraat</a> is SVP and Head of AI Acceleration at Truist and specializes in Agile AI at Enterprise Scale and Machine Learning and Deep Learning Innovation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f34c45d03e09" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Investing in your education — Is a Master’s Degree in Data Science “worth it?]]></title>
            <link>https://medium.com/@bjorn.austraat/investing-in-your-education-is-a-masters-degree-in-data-science-worth-it-ff228521b804?source=rss-6dd9a05cc909------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff228521b804</guid>
            <dc:creator><![CDATA[Bjorn Austraat]]></dc:creator>
            <pubDate>Wed, 09 Sep 2020 21:44:50 GMT</pubDate>
            <atom:updated>2020-10-19T12:36:11.208Z</atom:updated>
            <content:encoded><![CDATA[<h3>Investing in your education — Is a Master’s Degree in Data Science “worth it?”</h3><p><em>Article completed as part of the </em><a href="https://www.udacity.com/course/data-scientist-nanodegree--nd025"><em>Udacity Data Science Nanodegree</em></a><em> program October 2020</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*edZdjsLR4n8XYqyuyEbLKw.jpeg" /><figcaption>Will an advanced degree pay off big for data scientists?</figcaption></figure><p>Traditional advanced college degrees come with increasingly exorbitant costs. Even after the current Covid-19 pandemic abates and traditional in-person, in-classroom instruction resumes, many individuals who may have completed some or all of an undergrad degree will have to confront the question whether taking on potentially five- or six-figure debt to add a Master’s degree provides the appropriate long-term payoff.</p><p>I will use the <a href="https://insights.stackoverflow.com/survey/2020">2020 Stack Overflow survey data</a> as a starting point to explore the question of whether an aspiring data scientist is better off <strong>from a highly simplified cash flow / net worth perspective</strong> if they complete some or all of a Bachelor’s or Associate’s Degree and then either</p><ol><li>enroll in a relatively shorter and much more inexpensive bootcamp / self-guided / online-first education or</li><li>pursue a Master’s Degree with an expected higher starting salary but corresponding debt burden and deferred earnings potential.</li></ol><h3>When your (free) data is just slightly less than perfect…</h3><p>The Stack Overflow survey data is a treasure trove of global developer trends, preferences, salary figures, and more. Best of all, it’s a completely free resource that can be downloaded <a href="https://insights.stackoverflow.com/survey/">here</a>. As is true for all data sets, it’s important to inspect the underlying features and data points for logical consistency and relevance before basing any kind of analysis or machine learning process on these inputs.</p><p>For this exploration, I was most interested in starting salaries for individuals in the U.S. who are not employed in academia, had listed data science as (at least one) specialization, and who were early in their career. A first glance at the data revealed a large number of ‘interesting’ outliers when charting annual compensation against years of coding experience.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*RHgEmkxbBEEZfUhuQDeNlQ.png" /><figcaption>Noise!</figcaption></figure><p>Clearly, a salary of $2M with zero or 2 years of coding experience is improbable. There were many implausibly large salary figures throughout the spectrum and no clear correlation between experience and income at all! The survey methodology employed by Stack Overflow allows participants to specify salary based on varying intervals (week/month/year) and at least some of the noise may be due to respondents picking an incorrect multiplier (e.g. $30,000 per week rather than per year).</p><h3>Clean up on Aisle Data</h3><p>A deeper exploration of the distributions using a box plot diagram showed that the data did contain valuable signals but needed to be ‘de-noised’ first.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*orrZrGltEvFFTEAIPNBXGA.png" /><figcaption>A glimmer of hope</figcaption></figure><p>The tiny box plot at the very bottom shows the beginnings of a meaningful correlation between experience and income but required ‘quantile’ clean up first. Using the dataframe ‘quantile’ method, I pruned the top and bottom 10% of the survey values to get a more realistic picture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*Zu70pUqWKs-0cBBBKEu1WQ.png" /><figcaption>Much better!</figcaption></figure><p>Voila! Much more meaningful distribution and correlation with salary progression and widening of salary bands as people progress through their careers.</p><p>I next zoomed in on the two cohorts — some college or completed B.A./A.A. plus prospective online education = Mavericks vs. Master’s Degree = Traditional—to get a clearer picture of early-career median salaries.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*p00tX_SaLa-zM2y1ckcUqA.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*OSH_IbP1jo2LGfCOw2KCOA.png" /></figure><p>Based on the Stack Overflow survey data as well as other public domain data sources (<a href="http://dice.com/">Dice.com</a>, <a href="http://payscale.com/">Payscale.com</a>), I ultimately decided to compare a hypothetical Maverick with a Traditionalist using the following parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RO5fOIk2nDSyNPRGt_6uWg.png" /></figure><p>First, the Mavericks. They get to generate income very quickly and — assuming that they start putting money aside early in their career—can take advantage of compounding returns a full 1.5 years before their Traditionalist brethren.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*EWWZyM8WKomB8oTyV1XlUw.png" /></figure><p>Traditionalists (M.A.) start later, at a higher salary level, but also have to work off a significantly larger debt burden, which impairs their overall (simplified) net worth.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*eERssWwlmXEQJruD2iti9w.png" /></figure><p>Comparing the two cohorts side by side validates the intuition that additional education provides returns in the long run.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*00RkSMrgZEUL5fzTXllbew.png" /></figure><p>In conclusion, this highly simplified exploration provides the following insights:</p><ul><li>Traditionalists have a minor net worth advantage at the end of the 20 year observation period that will grow over time</li><li>Traditionalists have a pronounced salary advantage that will continue to grow over time</li><li>Starting to earn sooner (even at a lower level) pays off <strong>if </strong>you save some of your income in a compounding asset</li><li>The power of early compounding returns for Mavericks is significant and the large student loan debt for Traditionalists means that their net worth doesn’t catch up to Mavericks until 2037</li></ul><p>Naturally, there are many potential different outcomes, e.g. if annual salary increases on average are e.g. 6% for traditionalists vs. 5% for Mavericks, or if the savings rate were slightly different between cohorts, etc. etc.</p><p>So, to return to the question raised at the beginning of this post, the answer is “definitely maybe, but it depends!”</p><p>If you would like to run your own experiments with different parameters, feel free to download the data, Jupyter notebook, and Python source code from <a href="https://github.com/NYCDS2020/DataScienceBlogPost/">Github here</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff228521b804" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>