Improving Xbox Search Relevance by Click Likelihood Labeling

摘要

From the original game console, the Xbox has rapidly evolved into a comprehensive entertainment platform where tens of millions of users could not only play video games but also watch movies and TVs, listen music and enjoy Apps. Therefore, building a cross media ranker to provide relevant and personalized search results for Xbox users has become an interesting and imperative task. In this paper, we present our recent progress on improving Xbox's cross media ranker by mining massive click log data and generating multi-class relevance labels. Our experimental results have shown that incorporating the click likelihoods into the label generation yields better click-performance and meanwhile maintains comparable NDCG values, as compared to solely using the human labels generated by a small number of human judges.