Query Clustering for Learning to Rank Models on Web Search

摘要

Hundreds of features have been applied for ranking on nowadays Web search and it is unpractical to construct ranking functions through fusing these features with manually tuned parameters. So learning to rank, an interdisciplinary field of information retrieval and machine learning, has attracted increasing attention. However, the state-of-the-art learning to rank approaches try to find one general ranking function without concerning about the differences among the queries, which is not applicable in the diverse Web search scenario. In this paper, we address the issue of query representation and clustering for learning to rank. Queries are modeled as mean feature vectors and clustered through the largest connected sub-graph algorithm. Different ranking functions have been learned for each cluster. In the online search, the optimized function is selected for the new coming query according to its best fitted cluster. Two state-of-the-art learning to rank models namely Ranking SVM and ListMLE are studied in the comparative experiments. Results show that the proposed query-clustering based learning to rank approach makes significantly better or comparable performance.