Publications

SwiftXML: Extreme Multi-label Learning with Label Features for Warm-start Tagging, Ranking & Recommendation
Yashoteja Prabhu, Anil Kag, Shilpa Gopinath, Kunal Dahiya, Shrutendra Harsola, Rahul Agrawal, Manik Varma
accepted to WSDM 2018. [ArXiv]
Previous work on Extreme Classification had focused only on user features while completely ignoring label information. SwiftXML uses both user and label information to learn a partitioning leading to a classifier similar in scale to PfastreXML but better in performance. This results in 10% higher click-through rate and 31% lower bounce rate compared to existing algorithms in Bing Ads.

Parabel: Partitioned Label Trees for Extreme Classification with Application to Dynamic Search Advertising
Yashoteja Prabhu, Anil Kag, Shrutendra Harsola, Rahul Agrawal, Manik Varma
accepted to WWW (also known as TheWebConf) 2018. [ArXiv]
This work focuses on learning a hierarchy of coarse to fine label classifiers, each trained on a small subset of data points. This improves the precision@1 by 5% and resulted in 10x smaller model size in comparison to PfastreXML. This achieves 20% higher click-through rate and 32% lower bounce rate compared to existing algorithms in Bing Dynamic Search Ads.