<?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 Abishek Sainath Madduri on Medium]]></title>
        <description><![CDATA[Stories by Abishek Sainath Madduri on Medium]]></description>
        <link>https://medium.com/@abisheksm?source=rss-48398c154bd4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*W2E83T2p5Ff1cNlO.</url>
            <title>Stories by Abishek Sainath Madduri on Medium</title>
            <link>https://medium.com/@abisheksm?source=rss-48398c154bd4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 18:28:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@abisheksm/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[Predicting Spending Propensity]]></title>
            <link>https://medium.com/@abisheksm/predicting-spending-propensity-b5bd2bb23dfa?source=rss-48398c154bd4------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5bd2bb23dfa</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Abishek Sainath Madduri]]></dc:creator>
            <pubDate>Wed, 08 Feb 2017 20:04:26 GMT</pubDate>
            <atom:updated>2017-03-01T23:37:19.339Z</atom:updated>
            <content:encoded><![CDATA[<p>I worked on a consulting project for the company Hive(<a href="http://hive.co/">Hive.co</a>) to predict spending propensity of users based on app data and user data at Insight data science program(<a href="http://www.insightdatascience.com/">www.insightdatascience.com</a>)</p><p><strong>What is Hive?</strong></p><p>Hive is a platform for bands, labels, and brands to better understand their fan base. Hive collects personal information about fans, while tracking their engagement with respect to their clients. Some of this tracking is accomplished using users’ various social media accounts. One set of products within Hive’s platform are a set of promotional tools that allow their clients to create contests, fan club signup pages, and free online downloads, all of which users sign up for. Hive also tracks fans by analyzing Facebook comments, Instagram tags, Twitter retweets and other data related to a brands’ social media pages.</p><p><strong>Problem Addressed:</strong></p><p>The question that Hive was interested in was, given all this information about fans of a band, could we predict how likely someone is to spend money on a product like music or concert tickets or merchandise? They wanted a python module that would integrate with their existing pipeline to further enrich their clients understanding of their fans.</p><p><strong>Relevance of this problem:</strong></p><p>Social media data has not been used to predict spending propensity before and Hive has a lot of information on users that they could be mining for information. A good prediction pipeline also provides information to Hive’s clients about how successful their advertising is on social media and lets them justify their expenditure and adjust their social media strategy. A good prediction pipeline will result in more business for Hive. Currently they don’t do any prediction analysis and only target users if they have already spent money on a product.</p><p>Hive sent me some user data for a particular client of theirs and they were interested in seeing if I could predict spending propensity of fans that client. I got about ~150,000 users of which ~2000 users had spent money on products from this client. There was a huge imbalance in this data which I was training for spending propensity for : ~2000 users against ~148,000 users who were yet to spend money.</p><p><strong>Feature Engineering and Algorithm Analysis:</strong></p><p>My first step was to engineer features from the data given to me. For each social media platform, they tracked features like posts liked, retweeted, tagged etc. For example, for Facebook they tracked page likes, comments and tags. I computed a score for each of the social media features by getting an aggregate score representative of each action on a particular social media platform according to a certain weight associated with the time and action type. The weights were decided based on a conversation with Hive about how to compute this.</p><p>I implemented a few algorithms and present the top two based on classification accuracy. I performed a 10-fold cross validation procedure by randomly splitting the training and testing data and also adusted the threshold to get a good performance score. I dealt with the class imbalance problem by oversampling the underrepresented class. I also ensured that in the training data there was a 50–50 split of positive and negative examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/736/1*D1cE4aOSZ_rkwDSUPTIhgA.png" /></figure><p>The decision about which algorithm to use was based on the fact that while targeting users for spending propensity, the false negative rate is more important than false positive rate. Hive cares more about not missing a user who might spend money than sending notifications to someone who might not be interested in the product. For example, in this comparison — missing 4% of users from 90 to 94 translates to missing out on about $3000 for this client based on average money spent. On the other hand, precision scores matter also because sending too many notifications and targeting people might put them off Hive and even harm the reputation of the brand resulting in fans disengaging from Hive. I also looked at PCA to see if there were one or two components that explained all of the variation. This was not the case. I needed about 6 components to explain 80% variation in the data.</p><p>Based upon conversations with Hive, given the number of clients they have, they indicated that if this analysis scales up to their remaining clients, they expect to make at least 500,000$ on revenue over the next year.</p><p>I didn’t want to stop here and went a bit further to analyze which features were most important for the prediction of spending using recursive feature elimination of Logistic Regression based predictions —</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/898/1*LUTLc6b4hWTgdLYlaNz-Bg.png" /><figcaption>Feature Ranking</figcaption></figure><p>It turns out that of Hive’s promotional tools, downloading a song was the top predictor for buying and the top social media feature was SoundCloud. While one has to be careful attributing correlation to causation, this could be an indicator that SoundCloud activity is boosting up sales. To know this for sure, I would like to perform further analysis on the purchase data types which were not currently made available to me. Another insight that I was able to provide was that, among social media activity, activity on this client’s Soundcloud page was a good indicator for converting fans to paying customers. This helps Hive adjust social media strategy for their clients based on these predictors which is another avenue for making money for them.</p><p>The slides can be found at: <a href="http://bit.ly/HiveProjectSlides">bit.ly/HiveProjectSlides</a></p><h3><strong>About Me</strong></h3><p>I am currently finishing my PhD in Systems Biology at Weill Cornell Medical School. My research is about building gene regulatory networks in cancer to identify drivers of cancer and cancer drug targets. I have experience with working with large data sets such as the cancer genome data sets from TCGA which have about 20,000 genes and about 500 samples on average.</p><p>I completed a Masters in Computer Science where I wrote constructive proofs for schedulers and worked with Linux kernel scheduling algorithms.</p><p>My undergraduate degree was in Computer Science and Engineering where my area of research was in modeling schedulers for high performance clusters through a partial differential equation.</p><p>Linkedin : <a href="http://www.linkedin.com/in/abishek-sm">linkedin.com/in/abishek-sm</a></p><p>GitHub: <a href="http://github.com/abishek-sm">github.com/abishek-sm</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5bd2bb23dfa" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>