Community Clickthrough Model and Community Refinement for Query Recommendation

Abstract

Search engines have experienced significant improvement over the last two decades in terms of accuracy, coverage, functionality and speed. To help users with exploring relevant information space around their queries, some search engines also return related queries and concepts in addition to the matched web content. For example, Google recently announced the Knowledge Graph and celebrated it as a critical first step towards the next generation of search. Querying “Amadeus Mozart”, Google returns not only conventional results but also a number of related searches such as “amadeus mozart movie”, “amadeus mozart biography”, “amadeus mozart music”, etc., as well as links to various concepts (not web pages) about Mozart such as his birth date/place, his parents and compositions, etc. Clicking on those links, say, his birthplace, will lead to another set of links about Salzburg (under the embedded query of ”salzburg Austria”. Following the vision of knowledge graph, this project investigates techniques for dynamic query recommendations. We propose a community-based query recommendation approach by identifying the user's intention and interests in the information exploratory search process through the wisdom and prior experiences of similar users. in order to make relevant recommendations to users, we propose a 4-partite clickthrough model built on users, queries, documents and concepts and develop co-clustering algorithms to group objects under each of the four node types into clusters. To produce recommendations catering to users’ long-term and current interests, we propose to study methods that classify user interests into |ong—term, temporary and current interests and address the challenges in incorporating them into the clickthrough model. The user communities produced by clustering the clickthrough graph can be used to make both (i) high-quality recommendations for users to zoom into specified aspects of their queries and (ii) diversified recommendations for them to pan across different aspects of their queries into potentially distant but interesting topics. To address the problem of groupization, i.e., recommendations derived from a user’s communities may still be irrelevant to the user’s specific query, we propose to study different ways to refine user communities to ensure the relevance of the recommendations. We explore two possible dimensions for refinement: users who have conducted search that overlap with the user's current interests and users who have been very selective in their clickthroughs. Finally, we will conduct performance evaluation, user study and prototyping to demonstrate the success of the project.