Ankit Patil

Paul Anderson photo

Staff Software Engineer

Ankit is a Staff Software Engineer at Stubhub Inc with 6+ years of experience in Information Retrieval, Machine Learning and Software development. He helped StubHub to bootstrap query understanding, entity linking and relevance ranking. Prior to StubHub, he was with Telenav Inc and worked on compiling knowledge dictionary for address and POI search for the in-car navigation system.

Ankit Patil is speaking at the following session/s

Bootstrapping LETOR for suggestions

Thursday | 3:20PM - 4:00PM |

Search suggestions are the most frequently used functionality at StubHub with 12 Million+ average hits on a normal day. At StubHub, expectations for Search suggestions change from day to day. It is very challenging to use human ranked data set since data ranked on a day is not relevant after a few days. In this session, an approach to bootstrap machine learning based ranking for the most heavily used API is discussed. Baseline method uses sales as criteria for ranking. A logistic regression model is introduced where the weights were determined heuristically. This model performs better than the baseline model. The click-through data obtained from this model is used to train a more sophisticated neural net model to gain higher accuracy. A two-pass system that relies on Solr to obtain a candidate set of documents in the first pass and re-ranks the documents based on finer parameters in the second pass is being used.

Attendee Takeaway
Learn how to get started with LETOR for their existing system. Obtaining good training data is very important for any ML system and is the bottleneck for many organization to get started. We encourage and guide them to bootstrap ML-based ranking system with minimal efforts and risks.

Intended Audience
Mid level to senior Software Engineers, Engineering Managers, Data Scientists, Product Managers who want to introduce LETOR in their system. It would explain an intermediate, easy and low risk state which then eventually be transformed into full fledged LETOR using widely popular XGBoost or Neural network models. Some basic knowledge of Information retrieval and Machine Learning is required.

All Levels

Additional speakers