Effective Query Recommendation with Medoid-based Clustering using a Combination of Query, Click and Result Features

Abstract

Query recommendation is now an inseparable part of web search engines. The goal of query recommendation is to help users find their intended information by suggesting similar queries that better reflect their information needs. The existing approaches often consider the similarity between queries from one aspect (e.g., similarity with respect to query text or search result) and do not take into account different lexical, syntactic and semantic templates exist in relevant queries. In this paper, we propose a novel query recommendation method that uses a comprehensive set of features to find similar queries. We combine query text and search result features with bipartite graph modeling of user clicks to measure the similarity between queries. Our method is composed of two separate offline (training) and online (test) phases. In the offline phase, it employs an efficient k-medoids algorithm to cluster queries with a tolerable processing and memory overhead. In the online phase, we devise a randomized nearest neighbor algorithm for identifying most similar queries with a low response-time. Our evaluation results on two separate datasets from AOL and Parsijoo search engines show the superiority of the proposed method in improving the precision of query recommendation, e.g., by more than 20% in terms of p@10, compared with some well-known algorithms.