ABSTRACT:

Extreme Classification (XC) is a new paradigm of machine learning for solving multi-label problems with millions of categories. XC aims to provide practical solutions to large-scale recommendation and ranking problems. For example, we could reformulate search advertising as the problem of predicting the subset of search queries that an ad is most relevant to and solve it using XC algorithms. Such algorithms need to be highly efficient in order to learn accurate models from massive amounts of historical click logs and perform predictions in just a few milliseconds per ad. Moreover, the definition of accuracy itself needs to be redefined for such applications to account for extreme label imbalances & facilitate post-processing steps such as label filtering and re-ranking.

This talk presents some new XC techniques from our group which have led to significant impact in Bing advertising. Our novel Parabel and XReg algorithms achieve state-of-the-art prediction accuracies while maintaining low training and prediction complexities that scale only logarithmically with the number of categories. We also propose new evaluation metrics which are more suitable for the requirements of real-world ranking problems. Beyond advertising, we also apply our techniques to large-scale document tagging, item-to-item recommendation on Amazon, user-to-movie recommendation, etc. and observe consistent improvements in prediction qualities.

BIO:

Yashoteja Prabhu is a post-doctoral researcher at Microsoft Research India. He finished his PhD work at Indian Institute of Technology (IIT) Delhi under the supervision of Dr. Manik Varma. During PhD, he was awarded the TCS research fellowship. Prior to joining PhD, he spent two years at Microsoft Research India, working in the area of machine learning. He obtained his Bachelors degree in Computer Science from Indian Institute of Technology (IIT) Bombay. His current research interests fall under the broad theme of large-scale machine learning, covering the research areas of Extreme Classification, Ranking, Recommender Systems and Computational Advertising. He has authored several papers in top-tier computer science conferences such as WWW, WSDM and KDD.  More details about him can be found at www.cse.iitd.ac.in/~yashoteja.

Hosted by Professor Cho-Jui Hsieh