layout: default

title: Resume

Aaron Beppu


* 2005-2008 University of California Berkeley
BA in Cognitive Science, concentration in computational modeling
(Highest Honors, Departmental Citation, 3.8 GPA)


* January 2011 - Present
Software Engineer in Search * Designed and implemented system to improve search ranking by “learning” from repeated successes and failures. Offline clickstream analysis (Cascading+Hadoop) to discover the product features most important for a given query, coupled with online construction of augmented Solr queries favoring those features. Showed 20% improvement in conversion rate in A/B tests. * Analytics and dashboarding to compare search metrics (e.g. GMS per search, click rate, etc) between A/B test conditions. Rendered graphs in-browser using protovis/js. Metrics computed offline using Cascading/Hadoop. * Data-driven stemming. Hadoop analysis of clickstream data to discover words which our previous stemmer inappropriately reduced to the same form (as in “steel” and “Steelers”). Implemented new stemmer which keeps these words distinct, and integrated it into our analysis chain. * Autosuggestion of place names in a manner sensitive to population and distance from user's location. * July 2008 - December 2010 (a wholly-owned subsidiary of Amazon)
Software Development Engineer in Search Analytics
* Agglomerative clustering to find sets of similar queries and products based on user behavior data
* Generative models of user click behavior (and approximate Bayesian inference) to compensate for “position-bias” in ordered search results * Distributed system for orchestrating complex analytics work-flows
* Large scale ongoing analytics of search related user behavior using Hadoop
* Methods and tools to evaluate quality of spelling suggestions via Mechanical Turk
* Metrics to gauge proportion of Amazon's traffic, clicks, product views, purchases, and purchase dollar attributable to search
* Finding, summarizing and reporting sequences of user actions which strongly suggest that the user is having a poor search experience

Talks and Publications

* Presented “Data Mining for Product Search Ranking” at Hadoop World 2011. * Beppu, A., Griffiths, T. L. (2009). Iterated learning and the cultural ratchet. Proceedings of the 31st Annual Conference of the Cognitive Science Society. * The above paper describes cases when a sequence of communicating Bayesian agents will have the same asymptotic consistency as a single learner, and when they will not. The predictions of our model are validated using a lab-based learning experiment with human participants.