<?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[The Minds of MediaMath - Medium]]></title>
        <description><![CDATA[Views and news from the heart of adtech - Medium]]></description>
        <link>https://medium.com/mediamath?source=rss----9af5f972a7e7---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>The Minds of MediaMath - Medium</title>
            <link>https://medium.com/mediamath?source=rss----9af5f972a7e7---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 19:36:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/mediamath" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Machine Learning Demystified, Part 3: Models]]></title>
            <link>https://medium.com/mediamath/machine-learning-demystified-part-3-models-27823577bba0?source=rss----9af5f972a7e7---4</link>
            <guid isPermaLink="false">https://medium.com/p/27823577bba0</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[advertising]]></category>
            <dc:creator><![CDATA[prasad chalasani]]></dc:creator>
            <pubDate>Mon, 12 Sep 2016 16:55:48 GMT</pubDate>
            <atom:updated>2016-09-16T22:42:53.296Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*WhlJUiUWHsvCXE5S9qqfYQ.jpeg" /></figure><p>In <a href="https://medium.com/mediamath/machine-learning-without-tears-part-2-61f02552f165#.yieljykdl">Part 2</a> of our series, the ML novice realized that <strong>generalization </strong>is a key ingredient of a true ML algorithm. In today’s post we continue the conversation (the novice is in <strong><em>bold italics</em></strong>) to elaborate on generalization and take a first peek at how one might design an algorithm that generalizes well.</p><p>Last time we started talking about how one would design an ML algorithm to predict the click-through-rate (CTR) on an ad impression, and you designed two extreme types of algorithms. The first one was the <em>constant algorithm</em>, which simply records the <em>overall historical click-rate a</em>, and predicts <em>a </em>for every new ad opportunity.</p><p><strong><em>This was fine for overall accuracy but would do poorly on individual ad impressions, because its predictions are not differentiated based on the attributes of an impression, and CTRs could be very different for different types of impressions.</em></strong></p><p>Exactly, and the other extreme was the <em>memorization algorithm: </em>it records the historical CTR <em>for every combination of feature-values, </em>and for a new ad impression with feature-value-combination <em>x </em>(e.g., browser = “Safari”, gender = “male”, age = 32, city = “NYC”, ISP = “Verizon”, dayOfWeek = “Sunday”), it looks up the historical CTR <em>a(x) </em>(if any) for <em>x, </em>and outputs <em>a(x) </em>as its prediction.</p><p><strong><em>And we saw that this algorithm would be highly impractical and it would not generalize well to new ad impressions.</em></strong></p><p>Yes, it would not generalize well for two reasons: (a) the specific feature-value-combination <em>x </em>may not have been seen before, and (b) even if <em>x </em>occurred in the training data, it may have occurred too infrequently to reliably estimate CTR for cases resembling <em>x.</em></p><p><strong><em>So how would we design an algorithm that generalizes well?</em></strong></p><p>Let’s do a thought experiment: what do <em>you </em>do when you generalize in daily situations?</p><p><strong><em>I start from specific facts or observations and formulate a general concept.</em></strong></p><p>How, exactly, do you do that?</p><p><strong><em>I abstract common aspects of the specific observations, so that I can make a broader, more universal statement. I build a mental model that’s consistent with my observations.</em></strong></p><p>In terms of detail, would you say that your <em>mental model</em> is relatively <em>simple</em> or <em>complex</em>?</p><p><strong><em>Simple, definitely. I abstract away the non-essential details.</em></strong></p><p>Great, you’ve just identified some of the characteristics of a good generalization: a relatively <strong><em>simple</em></strong><em>, abstract, less detailed, model that is </em><strong><em>consistent</em></strong><em> with (or </em><strong><em>fits, </em></strong><em>or </em><strong><em>explains)</em></strong><em> the observations , and is more broadly applicable beyond the cases you have seen.</em></p><p><strong><em>That’s how humans think, but how does this help us design a computer algorithm that generalizes well?</em></strong></p><p>This helps us in a couple of ways. Firstly, it gives us a framework for designing an ML algorithm. Just like humans build mental models based on their observations, an ML algorithm should <em>ingest</em> training data and <em>output </em>a <strong>model</strong> that <em>fits, </em>or <em>explains </em>the training data well.</p><p><strong><em>What does a “model” mean specifically?</em></strong></p><p>A model is the mathematical analog to the human idea of a “concept” or “mental model”; it’s the mathematical formalization of what we have been informally calling “rules” until now. A model is essentially a <strong>function </strong>that<strong> </strong>takes as input the characteristics (i.e. features) of a case (i.e. example), and outputs a <em>classification</em> (e.g., “cat” or “not cat”) or <em>score</em> (e.g. click-probability).</p><p><strong><em>Wow, so a model is essentially a program, and you’re saying that an ML program is producing another program?</em></strong></p><p>Sure, that’s a good way to look at it, if you think of a function as a program that takes an input and produces an output.</p><p><strong><em>We seem to be back in voodoo-land: a program that ingests training data, and spits out a program…</em></strong></p><p>Well the ML algorithm does not magically output code for a model; instead the ML algorithm designer usually restricts the models to a certain <strong>class of models </strong><em>M</em><strong><em> </em></strong>that have a common structure or form, and the models in the class only differ by certain <strong>parameters </strong><em>p. </em>Think of the class of models <em>M </em>as a generic template, and a specific model from the class is selected by specifying the values of the parameters <em>p. </em>Once restricted to a certain class of models, the ML algorithm only needs to <em>find the values of the parameters</em> <em>p </em>such that the specific model with these parameter-values<em> </em>fits the training data well<em>.</em></p><p>For example suppose we’re trying to predict the sale price <em>y </em>of a home in a certain locality based only on the square footage <em>x</em>. If we believe home prices are roughly proportional to their square footage, a simple class of models for this might be the class of <strong>linear models</strong>, i.e., the home price <em>y </em>is modeled as <em>a*x. </em>The reason these models are called linear is that, for a fixed value of <em>a, </em>a plot of <em>x</em> versus <em>y </em>would be a straight line.</p><p>Notice that <em>a*x </em>is a generic template for a linear model, and a value for <em>a </em>needs to be specified to get a specific linear model, which would then be used to predict the sale price for a given square-footage <em>x. </em>The ML algorithm would be trained on examples of pairs <em>(x,y)</em> of square-footage and home-price values, and it needs to find the value of the parameter <em>a </em>such that the model <em>a*x</em> “best fits” the training data. I hope this does not sound so mysterious any more?</p><p><strong><em>It certainly helps! Instead of magically spitting out code for some model, an ML algorithm is actually “merely” finding parameters p for a specific class of models.</em></strong></p><p>Right…</p><p><strong><em>But how do we know a model is “correct” or is the “true” one?</em></strong></p><p>Remember that a model expresses a relationship between features and response (e.g. classification or score). All non-trivial real-world phenomena are governed by some “true” underlying model (think of this as the <strong><em>signal</em></strong><em>, </em>or the<em> essence</em>, of the relationship between the variables involved) combined with random <strong>noise</strong>. The noise could result from errors in measurement or inherent random variation in the relationship. However an ML algorithm does not need to discover this “true” underlying model in order to be able to generalize well; it suffices to find a model that fits the training data adequately. As the statistician George Box famously said,</p><blockquote>All models are wrong, but some are useful,</blockquote><p>meaning that in any non-trivial domain, every model is at best an approximation when examined closely enough, but some models can be more useful than others.</p><p><strong><em>What kinds of models would be more useful?</em></strong></p><p>In the ML context, a useful model is one that generalizes well beyond the training data. This brings us to the second way in which thinking about how humans generalize helps, which is that it suggests a guiding principle in designing a true learning algorithm that has good generalization abilities. This principle is called <strong>Occam’s Razor:</strong></p><blockquote>Find the <strong>simplest</strong> model that explains the training data well. Such a model would be more likely to generalize well beyond the training examples.</blockquote><p><strong><em>And what is a “simple” model?</em></strong></p><p>In a given context or application-domain, a model is <em>more complex </em>if its structure or form is more complex and/or it has more parameters (the two are often related). For example when you’re predicting CTR, and you have thousands of features you could be looking at, but you’re able to <em>explain the training data “adequately”</em> using just 3 of them, then the Occam’s Razor principle says you should just use these 3 to formulate your CTR-prediction model. Or when predicting house prices as a function of square-footage, if a linear model <em>fits the training data adequately</em>, you should use this model (rather than a more complex model such as quadratic, etc) as it would generalize better to examples outside the training set. I’ve used italics when talking about fitting/explaining the training data because that’s something I haven’t explained yet.</p><p><strong><em>I see, so the constant algorithm I designed to predict CTR produces an extremely simple model that does not even fit the training data well, so there’s no hope it would generalize beyond the training data-set?</em></strong></p><p>Exactly, and on the other hand the <em>memorization</em> algorithm produces an extremely complex model: it records the historical CTR for each feature-value combination in the training data. When tested on any example (i.e. feature-value combination) that already occurred in the training data, this model would produce exactly the historical CTR observed for that example. In this sense this complex model <em>fits the training data perfectly. </em>When an overly complex model fits training data exceptionally well, we say the model is <strong>overfitting </strong>the training data. Such models would not generalize well beyond the training data-set.</p><p><strong><em>Interesting, so we want to find a relatively simple model that fits the training data “adequately”, but we don’t want to overdo it, i.e. we shouldn’t overly complicate our model in order to improve the fit to the training data. Is it obvious why complex models in general wouldn’t generalize well?</em></strong></p><p>It’s actually not obvious, but here’s an intuitive explanation. Remember we said that any real-world phenomenon is governed by some true model that represents the signal or essence of the phenomenon, <em>combined with random noise</em>. A complex model-class has so many degrees of freedom (either in its structure or number of parameters) that when a best-fitting model is found from this class, it overfits the training data: it fits the signal as well as the random noise. The random noise gets baked into the model and this makes the model perform poorly when applied to cases not in the training set.</p><p><strong><em>Can you give me an example where a simpler model generalizes better than a complex one?</em></strong></p><p>Let’s look at the problem of learning a model to predict home sale prices in a certain locality based on the square-footage. You’re given a training data-set of pairs <em>(x,y) </em>where <em>x </em>is the square-footage in thousands, and <em>y </em>is the sale price in millions of dollars:</p><blockquote>(2, 4.2), (3, 5.8), (4, 8.2), (5, 9.8), (6, 12.2), (15, 27.8),</blockquote><p>and your task is to find a model that “best” fits the data (you decide how to interpret “best”). Any ideas?</p><p><strong><em>I notice from the training data that when the square-footage x is an even number, the home price is (2*x + 0.2), and when it’s an odd number the home price is (2*x — 0.2).</em></strong></p><p>Great, your model fits the training data <em>perfectly, </em>but in order to do so you’ve made your model complex because you have two different expressions based on whether the square footage is even or odd. Can you think of a simpler model?</p><p><strong><em>I see what you mean — I notice that the home prices (in millions of dollars) in the training data-set are fairly close to twice the number of thousands of square feet, so my simpler model would be 2*x.</em></strong></p><p>This would be a simpler model, even though it does not exactly fit the training data: the home prices are not exactly <em>2*x </em>but close to it. We can think of the home price as being <em>2*x </em>plus some random noise (which could be positive or negative). It just so happens that in the specific training examples you’ve seen, the noise part in the even-square-footage examples is 0.2, and -0.2 for the others. However it could well be that for a new example with <em>x=7 </em>(an odd number)<em> </em>the price is exactly <em>2*x</em> or slightly above <em>2*x</em>, but if we were to use your previous complex model, we would be predicting <em>2*x — 0.2, </em>which would not be a good prediction.</p><p>The point is that in an effort to fit the training data very well (in your case perfectly), you made your model overly complex and latched on to a <em>spurious pattern </em>that is unlikely to hold beyond the training data-set. By definition the noise part is random and un-predictable, so any attempt to model it and predict would result in errors. Instead, the simpler model captures only the “signal” and is oblivious to the random noise.</p><p>Here, then, is the recipe for designing a true ML algorithm that generalizes well:</p><ul><li>Pick a relatively simple <em>class of models</em> <em>M, </em>parametrized by a set of <em>parameters</em> <em>p</em></li><li>Find values of the parameters <em>p </em>such that the model from <em>M </em>with these parameters <em>fits the training data optimally</em>. Denote this model as <em>M[p]. M[p] </em>is a function that maps a feature-value combination <em>x </em>to a response (e.g. classification or score).</li><li>Output <em>M[p] </em>as the predictive model for the task, i.e. for an example <em>x, </em>respond with <em>M[p](x) </em>as the predicted classification or score.</li></ul><p><strong><em>What does it mean for a model to fit or explain the training data, and how do I find the parameters that “optimally” fit the data?</em></strong></p><p>Informally, a model fits the training data well when it is <em>consistent</em> with the training data. This is analogous to how you form a mental model based on your observations: your mental model is consistent with your observations (at least if you’re observant enough and think rationally). Of course to implement a computer algorithm we need to make precise the notion of an “optimal fit”, and specify how the computer should go about finding model parameters that “best” fit the training data. These are great topics for a future post!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=27823577bba0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/mediamath/machine-learning-demystified-part-3-models-27823577bba0">Machine Learning Demystified, Part 3: Models</a> was originally published in <a href="https://medium.com/mediamath">The Minds of MediaMath</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning without Tears, Part 2]]></title>
            <link>https://medium.com/mediamath/machine-learning-without-tears-part-2-61f02552f165?source=rss----9af5f972a7e7---4</link>
            <guid isPermaLink="false">https://medium.com/p/61f02552f165</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[digital-marketing]]></category>
            <dc:creator><![CDATA[prasad chalasani]]></dc:creator>
            <pubDate>Mon, 22 Aug 2016 15:35:43 GMT</pubDate>
            <atom:updated>2016-08-23T12:37:37.753Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*u6_Sizm2Ow6yyKfxWs899Q.jpeg" /></figure><p>In the <a href="https://medium.com/mediamath/machine-learning-for-the-perplexed-part-1-f67e74f6fb75#.1q2968mk5">first post</a> of our non-technical ML intro series we discussed some general characteristics of ML tasks. In this post we take a first baby step towards understanding how learning algorithms work. We’ll continue the dialog between an ML expert and an ML-curious person.</p><p><strong><em>Ok I see that an ML program can improve its performance at some task after being trained on a sufficiently large amount of data, without explicit instructions given by a human. This sounds like magic! How does it work?</em></strong></p><p>Let’s start with an extremely simple example. Say you’re running an ad campaign for a certain type of running shoe on the NYTimes web-site. Every time a user visits the web-site, an<em> ad-serving opportunity</em> arises, and given the <strong>features</strong> of the ad-opportunity (such as time, user-demographics, location, browser-type, etc) you want to be able to predict the chance of the user clicking on the ad. You have access to <strong>training examples: </strong>the last 3 weeks of <em>historical</em> logs of features of ads served, and whether or not there was a click. Can you think of a way to write code to predict the click-rate using this training data?</p><p><strong><em>Let me see, I would write a program that looks at the trailing 3 weeks of historical logs, and if N is the total of ad exposures, and k is the number of those that resulted in clicks, then for any ad-opportunity it would predict a click probability of k/N.</em></strong></p><p>Great, and this would be an ML program! The program ingests historical data, and given any ad-serving opportunity, it outputs a click probability. If the historical (over the trailing 3-weeks) fraction of clicked ads changes over time, your program would change its prediction as well, so it’s adapting to changes in the data.</p><p><strong><em>Wow, that’s it? What’s all the fuss about Machine Learning then?</em></strong></p><p>Well this would be a very rudimentary learning algorithm at best: it would be accurate <em>in aggregate</em> over the whole population of ad-exposures. What if you want to improve the accuracy of your predictions for <em>individual</em> ad-opportunities?</p><p><strong><em>Why would I want to do that?</em></strong></p><p>Well if your goal is to show ads that are likely to elicit clicks, and you want to figure out how much you want to pay for showing an ad, the most important thing to predict is the click probability (or CTR, the click-through-rate) <em>for each specific ad opportunity</em>: you’ll want to pay more for higher CTR opportunities, and less for lower CTR opps.</p><p>Say you’re running your ad campaign in two cities: San Francisco and Minneapolis, with an equal number of exposures in each city. Suppose you found that overall, 3% of your ads result in clicks, and this is what you predict as the click-probability for <em>any</em> ad opportunity. However when you look more closely at the historical data, you realize that <em>all ad-opportunities are not the same:</em> You notice an interesting <em>pattern</em>, i.e. 5% of the ads shown to users in San Francisco are clicked, compared to only 1% of ads shown to users logging in from Minneapolis. Since there are an equal number of ads shown in the two cities, you’re observing an average click-rate of 3% overall, and …</p><p><strong><em>Oh ok, I know how to fix my program! I will put in a simple rule: if the ad opportunity is from San Francisco, predict 5%, and if it’s from Minneapolis, predict 1%. Sorry to interrupt you, I got excited…</em></strong></p><p>That’s ok… in fact you walked right into a trap I set up for you: you gave a perfect example of an <em>ad-hoc static rule:</em> You’re <em>hard-coding an instruction</em> in your program that leverages a specific pattern you found by <em>manually</em> slicing your data, so this would <em>not</em> be an ML program at all!</p><p><strong><em>So… what’s so bad about such a program?</em></strong></p><p>Several things: (a) this is just <em>one pattern</em> among many possible patterns that could exist in the data, and you just happened to find this one; (b) you discovered this pattern by <em>manually</em> slicing the data, which requires a lot of time, effort and cost; (c) the patterns can <em>change</em> over time, so a hard-coded rule may cease to be accurate at some point. On the other hand, a learning algorithm can find <em>many relevant patterns</em>, <em>automatically</em>, and can <em>adapt</em> over time.</p><p><strong><em>I thought I understood how a learning algorithm works, now I’m back to square one!</em></strong></p><p>You’re pretty close though. Instead of hard-coding a rule based on a specific pattern that you find manually, you write code to slice historical data by <em>all</em> features. Suppose there were just 2 features:<em>city</em> (the name of the city) and <em>IsWeekend </em>(1 if the opportunity is on a weekend, 0 otherwise). Do you see a way to improve your program so that it’s more general and avoids hard-coding a specific rule?</p><p><strong><em>Yes! I can write code to go through all combinations of values of these features in the historical data, and build a lookup table showing for each (city, IsWeekend) pair, what the historical click-through-rate was. Then when the program encounters a new ad-opportunity, it will know which city it’s from, and whether or not it’s a weekend, and so it can lookup the corresponding historical rate in the table, and output that as its prediction.</em></strong></p><p>Great, yes you could do that, but there are a few problems with this solution. What if there were 30 different features? Even if each feature has only 2 possible values, that is already 2³⁰ possible combinations of values, or more than a billion (and of course, the number of possible values of many of the features, such as cities, web-sites, etc could be a lot more than just two). It would be very time-consuming to group the historical data by these billions of combinations, our look-up table would be huge, and so it would be very slow to even make a prediction. The other problem is this: what happens when an ad opportunity arises from a <em>new city that the campaign had no prior data for? </em>Even if we set aside these two issues, your algorithm’s click-rate predictions would in fact most likely <em>not </em>be very accurate at all.</p><p><strong><em>Why would it not work well?</em></strong></p><p>Your algorithm has essentially <strong>memorized</strong><em> </em>the click-rates for all possible feature-combinations in the training data, so it would perform excellently if its performance is evaluated on the training data: the predicted click-rates would exactly match the historical rates. But predicting on <em>new</em> ad opportunities is a different matter; since there are 30 features, each with a multitude of possible values, it is highly likely that these new opportunities will have feature-combinations that were never seen before.</p><p>A more subtle point is that even if a feature-combination has occurred before, simply predicting the historical click-rate for that combination might be completely wrong: for example suppose there were just 3 ad-opportunities in the training data which had this feature-combination: (Browser = “safari”, IsWeekend = 1, Gender = “Male”, Age = 32, City = “San Francisco”, ISP = “Verizon”), and the ad was not clicked in all 3 cases. Now if your algorithm encounters a new opportunity with this exact feature-combination, it would predict a 0% click-rate. This would be accurate with respect to the historical data your algorithm was trained on, but if we were to test it on a realistic distribution of ad opportunities, the prediction would almost certainly not be accurate.</p><p>What went wrong here? Suppose the true click-rate for ads with the above feature-combination is 1%, then in a historical sample where just 3 such ad-opportunities are seen,<em> </em>it’s statistically very likely that we would see no clicks.</p><p><strong><em>But what could the learning algorithm do to avoid this problem? Surely it cannot do any better given the data it has seen?</em></strong></p><p>Actually it can. By examining the training data, it should be able to realize, for example, that the ISP and Browser features are not relevant to predicting clicks (for this specific campaign), and perhaps it finds that there are a 1000 training examples (i.e. ad-opportunity feature-combinations) that match the above example when ISP and Browser are ignored, and 12 of them had clicks, so it would predict a 1.2% click-rate.</p><p>So your algorithm, by memorizing the click-rates from the training data at a very low level of granularity, was “getting lost in the weeds” and was failing to <strong>generalize </strong>to new data. The ability to generalize is crucial to any useful ML algorithm, and indeed is a hallmark of intelligence, human or otherwise. For example think about how you learn to recognize cats: you don’t memorize how each cat looks and try to determine whether a new animal you encounter is a cat or not by matching it with your memory of a previously-seen cat. Instead, you learn the <strong>concept </strong>of a “cat”, and are able to generalize your ability to recognize cats beyond those that exactly match the ones you’ve seen.</p><p>In the next post we will delve into some ways to design true learning algorithms that generalize well.</p><p><strong><em>Ok, looking forward to that. Today I learned that generalization is fundamental to machine-learning. And I will memorize that!</em></strong></p><p>This post originally appeared on the <a href="http://blog.mediamath.com//?utm_source=Medium&amp;utm_medium=Social&amp;utm_campaign=&amp;utm_content=MachineLearningPart2">MediaMath blog</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61f02552f165" width="1" height="1" alt=""><hr><p><a href="https://medium.com/mediamath/machine-learning-without-tears-part-2-61f02552f165">Machine Learning without Tears, Part 2</a> was originally published in <a href="https://medium.com/mediamath">The Minds of MediaMath</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning for the Perplexed, Part 1]]></title>
            <link>https://medium.com/mediamath/machine-learning-for-the-perplexed-part-1-f67e74f6fb75?source=rss----9af5f972a7e7---4</link>
            <guid isPermaLink="false">https://medium.com/p/f67e74f6fb75</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[advertising]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[big-data]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[prasad chalasani]]></dc:creator>
            <pubDate>Fri, 19 Aug 2016 18:54:56 GMT</pubDate>
            <atom:updated>2016-09-14T21:06:39.719Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/688/1*6Q0KKeTR0lFc7t4UKoh9Sg.jpeg" /></figure><p>With the increasingly vast volumes of data generated by enterprises, relying on static rule-based decision systems is no longer competitive; instead, there is an unprecedented opportunity to optimize decisions, and adapt to changing conditions, by leveraging patterns in real-time and historical data.</p><p>The very size of the data however makes it impossible for humans to find these patterns, and this has lead to an explosion of industry interest in the field of Machine Learning, which is the science and practice of designing computer algorithms that, broadly speaking, find patterns in large volumes of data. ML is particularly important in digital marketing: understanding how to leverage vast amounts of data about digital audiences and the media they consume can be the difference between success and failure for the world’s largest brands. At MediaMath, where I am the SVP of Data Science, our vision is for every addressable interaction between a marketer and a consumer to be driven by ML optimization against all available, relevant data at that moment, to maximize long-term marketer business outcomes.</p><p>In this series of blog posts we will present a very basic, non-technical introduction to Machine Learning. In today’s post we start with a definition of ML in the form of a dialog between you and an ML expert. When we say “you”, we have in mind someone who is <em>not</em> an ML expert or practitioner, but someone who has heard about Machine Learning and is curious to know more.</p><p><strong><em>Can we start at the beginning? What is Machine Learning?</em></strong></p><p>Machine learning is the process by which a computer program improves its performance at a certain task with experience, without being given explicit instructions or rules on what to do.</p><p><strong><em>I see, so you’re saying the program is “learning” to improve its performance.</em></strong></p><p>Yes, and this is why ML is a branch of Artificial Intelligence, since learning is one of the fundamental aspects of intelligence.</p><p><strong><em>When you say “with experience,” what do you mean?</em></strong></p><p>As the program gains “practice” with the task, it gets better over time, much like how we humans learn to get better at tasks with experience. For example an ML program can learn to recognize pictures of cats when shown a sufficiently large number of examples of pictures of “cat” and “not cat”. Or an autonomous driving system learns to navigate roads after being <em>trained</em> by a human on a variety of types of roads. Or a Real-Time-Bidding system can learn to predict users’ propensity to convert (i.e. make a purchase) when exposed to an ad, after observing a large number of historical <em>examples</em> of situations (i.e. combinations of user, contextual, geo, time, site attributes) where users converted or not.</p><p><strong><em>You said “without being given explicit instructions.” Can you expand on that a bit?</em></strong></p><p>Yes that is a very important distinction between an ML program and a program with human-coded rules. As you can see from the above examples, an ML system in general needs to <em>respond </em>to a <em>huge</em> variety of possible situations: e.g., respond “cat” when shown a picture of a cat, or turn the steering wheel in the right direction in respond to the visual input of the road, or compute a probability of conversion when given a combination of features of an ad impression. The sheer variety of number of possible input pictures, or road-conditions, or impression-features is enormous. If we did not have an ML algorithm for these tasks we would need to <em>anticipate </em>all possible inputs and program explicit rules that we hope will be appropriate responses to those inputs.</p><p><strong><em>I still don’t understand why it’s hard to write explicit rules for these tasks. Humans are very good at recognizing cats, so why can’t humans write the rules to recognize a cat?</em></strong></p><p>That’s a great question. It’s true that humans excel at learning certain tasks, for example recognizing cats, or recognizing handwriting, or driving a car. But here’s the paradoxical thing — while we are great at these tasks, the process by which we accomplish these tasks cannot be boiled down to a set of rules, even if we’re allowed to write a huge number of rules. So these are examples of tasks where explicit rules are <em>impossible</em> to write.</p><p>On the other hand there are tasks at which humans are <em>not</em> even good at: for example trying to predict which types of users in what contexts will convert when exposed to ads. Marketing folks might have <em>intuition </em>about what conditions lead to more conversions, such as “users visiting my site on Sundays when it’s raining are 10% likely to buy my product”. The problem though is that these intuition-guided rules can be <em>wrong, </em>and <em>incomplete </em>(i.e. do not cover all possible scenarios). The only way to come up with the right rules is to pore through millions of examples of users converting or not, and extract patterns from these, which is precisely what an ML system can do. Such pattern extraction is beyond the capabilities of humans, even though they are great at certain other types of pattern extraction (such as visual or auditory).</p><p><strong><em>I see, so ML is useful in tasks where (a) a response is needed on a huge number of possible inputs, and (b) it’s impossible or impractical to hard-code rules that would perform reasonably well on most inputs. Are there examples where the number of possible inputs is huge, but it’s easy to write hard-coded rules?</em></strong></p><p>Sure: I’ll give you a number, can you tell if it’s even or odd? Now you wouldn’t need an ML program for that!</p><p>This post originally appeared on the <a href="http://blog.mediamath.com//?utm_source=Medium&amp;utm_medium=Social&amp;utm_campaign=&amp;utm_content=MachineLearningPart1">MediaMath blog</a>. In the <a href="https://medium.com/mediamath/machine-learning-without-tears-part-2-61f02552f165#.n0qrgr8cg">next post</a> we will take a first step towards understanding how ML algorithms actually work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f67e74f6fb75" width="1" height="1" alt=""><hr><p><a href="https://medium.com/mediamath/machine-learning-for-the-perplexed-part-1-f67e74f6fb75">Machine Learning for the Perplexed, Part 1</a> was originally published in <a href="https://medium.com/mediamath">The Minds of MediaMath</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[MediaMath’s Story: Our Founder Joe Zawadzki Explains]]></title>
            <link>https://medium.com/mediamath/mediamaths-story-our-founder-joe-zawadzki-explains-6dedc4f57481?source=rss----9af5f972a7e7---4</link>
            <guid isPermaLink="false">https://medium.com/p/6dedc4f57481</guid>
            <category><![CDATA[adtech]]></category>
            <category><![CDATA[programmatic]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[dsp]]></category>
            <category><![CDATA[careers]]></category>
            <dc:creator><![CDATA[MediaMath]]></dc:creator>
            <pubDate>Thu, 18 Aug 2016 20:00:14 GMT</pubDate>
            <atom:updated>2016-08-23T12:39:11.447Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*kaJP62yBlnAAcKGjHBry4A.jpeg" /></figure><p>MediaMath will celebrate its 10th birthday next year. A lot has changed since 2007, when Joe Zawadzki, Erich Wasserman and Greg Williams came together to start the industry’s first demand-side platform. From a shoebox office of a handful of employees to a global operation with 750 employees in 16 locations around the world, MediaMath has certainly grown into its own. But the early days weren’t easy, as our CEO Joe Z explans in this video on our founder’s story. <a href="http://www.mediamath.com/resource/mediamaths-story-our-founder-joe-zawadzki-explains/">Watch the clip</a> to hear about one fateful Thanksgiving in the early days of the business when he accidentally turned a client campaign upside down and the work it has taken to get MediaMath where it is today.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FTKH08Gwo_G8%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DTKH08Gwo_G8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FTKH08Gwo_G8%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bb9bce5e1169ebd78b2e5e4f63eea370/href">https://medium.com/media/bb9bce5e1169ebd78b2e5e4f63eea370/href</a></iframe><p>Read more on the <a href="http://blog.mediamath.com//?utm_source=Medium&amp;utm_medium=Social&amp;utm_campaign=&amp;utm_content=MediaMathsFoundersStory">MediaMath blog</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6dedc4f57481" width="1" height="1" alt=""><hr><p><a href="https://medium.com/mediamath/mediamaths-story-our-founder-joe-zawadzki-explains-6dedc4f57481">MediaMath’s Story: Our Founder Joe Zawadzki Explains</a> was originally published in <a href="https://medium.com/mediamath">The Minds of MediaMath</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Employee Spotlight: From Tech’s Infancy to Programmatic]]></title>
            <link>https://medium.com/mediamath/employee-spotlight-from-techs-infancy-to-programmatic-b4e59103f545?source=rss----9af5f972a7e7---4</link>
            <guid isPermaLink="false">https://medium.com/p/b4e59103f545</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[digital-marketing]]></category>
            <category><![CDATA[careers]]></category>
            <dc:creator><![CDATA[MediaMath]]></dc:creator>
            <pubDate>Thu, 18 Aug 2016 19:50:58 GMT</pubDate>
            <atom:updated>2016-08-23T12:41:20.662Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/693/1*qEr9nP3e0AenIRxcZ41UXQ.jpeg" /></figure><p>“I think I was lucky when I came out of school at a very exciting time when the internet was really nascent and the “Gold Rush” was just about to begin,” said Jen Gold, Director of Product Strategy for Helix at MediaMath. “It was about the geeking out of it and the passion and the excitement of getting on the internet for the first time and thinking, this is like a wild west of information and content!”</p><p>From joining the digital tech space in its infancy to witnessing the dot-com bust and everything in-between, Gold didn’t step foot on the tech turf until after graduation.</p><p>She jokingly said she started in a very high tech field — majoring in intellectual history at the University of Pennsylvania. “It was a combination of history, philosophy, political science and literature, as well as art. Not high tech at all! I initially really wanted to get into some form of media and advertising but really, digital wasn’t a big thing back then.”</p><p>Gold’s first job out of school was with the <em>New York Times</em>, working in the advertising division on sales planning.</p><p>“I stayed there for a year and a half and it really wasn’t a tech job but I was put on a task force to figure out what to do with the content online.”</p><p>Here, Gold got to work on what content should be put out on AOL, CompuServe or the actual world wide web, which was starting to become a content rich place.</p><p>“The second I got on the internet I was like this is what I want to be doing, I’ve got to get out of this old school media place and into digital, so I left and joined a very small 18-person website development company startup and have been in that realm for most of my career since then.”</p><p>The boom was really happening around this time in the late 90s — Netscape went public among others and companies could do no wrong. But even when the bubble burst, Gold stayed in tech.</p><p>“I joined a travel guide book company called <em>Rough Guides</em> where I managed all their digital strategy. I’ve always been really into travel and totally fascinated by traveling around the world.”</p><p><strong>Watch the video below!</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FdJAdWv2INQk%3Ffeature%3Doembed&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DdJAdWv2INQk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FdJAdWv2INQk%2Fhqdefault.jpg&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e6fe9734480872c40e2d6aa65754507a/href">https://medium.com/media/e6fe9734480872c40e2d6aa65754507a/href</a></iframe><p>This led to her next adventure.</p><p>In 2004, Gold and her partner decided to make a huge leap of faith and leave New York City for a tiny island in the Caribbean called Vieques, off the coast of Puerto Rico. They bought property there and spent a lot of time on renovation, turning it into a guest house. Gold was fortunate enough to keep her job with <em>Rough Guides</em>, working remotely while learning about the hospitality business. It was during this time she inadvertently became a marketer.</p><p>“It wasn’t on my radar but in marketing my own hospitality business, I found that I really liked it. I got into SEO, SEM, some paid display and started experimenting with a little mobile and social and I was hooked!”</p><p>Gold still owns the business, but moved back to the States working on product and marketing with<em>TripAdvisor</em> for a few years, where she then made her way to <em>MediaMath</em>. Throughout her interview process, Gold says she was excited about the opportunities here and the growth of programmatic, working on product strategy for MediaMath’s proprietary shared data asset.</p><p>“I synthesize what our product teams are building in to something that’s packaged and digestible, which our client services organization and engagement teams can easily take to market and explain to clients what it is we have and why they need it. I find that really rewarding.”</p><p>Reflecting back on her career thus far, what would Gold tell her younger self?</p><p>“I love the fact that I was able to get right into digital pretty quickly. I guess if I was to say anything to my recently college graduated self, I would just say buckle in, get ready for the ride and be ready for lots of change!”</p><p>This article originally appeared on the <a href="http://blog.mediamath.com//?utm_source=Medium&amp;utm_medium=Social&amp;utm_campaign=&amp;utm_content=FromTechsInfancytoProgrammatic">MediaMath blog</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4e59103f545" width="1" height="1" alt=""><hr><p><a href="https://medium.com/mediamath/employee-spotlight-from-techs-infancy-to-programmatic-b4e59103f545">Employee Spotlight: From Tech’s Infancy to Programmatic</a> was originally published in <a href="https://medium.com/mediamath">The Minds of MediaMath</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>