Combining Personalization and Groupization to Enhance Web Search

Abstract

To improve retrieval effectiveness, personalized search engines adjust the search results according to the user’s interest profile. Most of the existing work are based on either pure personalization or pure groupization. Personalization tries to produce results aligned with the user’s interests, whereas groupization aims to broaden the results using the interests of the user’s communities. In this paper, we propose to combine personalization and groupization to improve the retrieval effectiveness of a personalized search engine. We observe that recommendations derived from a user’s communities may be too broad to be relevant to a user’s specific query. Thus, we study different ways to refine user communities according to the user’s personal preferences to improve the relevance of the recommendations. We introduce online user community refinement to identify highly related users and use their preferences to train a community-based personalized ranking function to improve the search results. To produce the user communities, we propose the Community Clickthrough Model (CCM) to model search engine clickthroughs as a tripartite graph involving users, queries and concepts embodied in the clicked pages, and develop the Community-based Agglomerative-Divisive Clustering (CADC) algorithm for clustering the CCM graph into groups of similar users, queries and concepts to support community-based personalization. Experimental results show that a refined user community that only uses focused users’ recommendations can significantly improve nDCG by 54% comparing to the baseline method. We also confirm that CADC can efficiently cluster CCM to enhance a personalized search engine.