Query Dependent Ranking based on PCA-based Query Representation

Abstract

Ranking is a crucial part of information retrieval. Queries describe the users’ search intent and therefore they play an essential role in the context of ranking for information retrieval. The diverse feature impacts on ranking relevance with respect to different queries. This paper tends to consider query difference in learning ranking function by clustering the queries where each query cluster represents a group of queries which have the similar set of important features for measuring ranking relevance. The success of clustering usually depends on the representation of the data. The query features are generated based on the ranking features values of querydocument pair and Principal Component Analysis (PCA) is used to construct the representation of query. To cluster the queries, bisecting k-means clustering algorithm is used. RankSVM algorithm is used for model construction.