Query recommendation based on irrelevant feedback analysis

Abstract

Similarity computation among queries is a central step of query recommendation based on click information in search log. In this step, weights of clicked URLs or clicked document terms, which may have a large influence on similarity computation results, are mostly counted based on co-occurrence. However, counting weights based on co-occurrence are unusually disturbed by irrelevant feedbacks in search log, which may decrease the precision of query similarity computation. This paper proposes a method that computes similarity among queries based on "Query - Clicked Sequence" model, which counts weight of clicked document term by density of documents containing this term on clicked sequence, and filters content of irrelevant documents during similarity computation. A series of experiment results show that this method can precisely count the weights of terms, and increase the precision of query similarity computation, accordingly increase the precision of query recommendation.