Personalized document ranking: Exploiting evidence from multiple user interests for profiling and retrieval

The goal of personalization in information retrieval is to tailor the search engine results to the specific goals, preferences and general interests of the users. We propose a novel model for both user profiling and document ranking that consider the user interests as sources of evidence in order to tune the accuracy of the documents returned in response to the user query. User profiling is performed by managing the user search history using statistical based operators in order to highlight the user short-term interests seen as surrogates for building the long-term ones. The document ranking model's foundation comes from influence diagrams which are extension of Bayesian graphs, dedicated to decision-making problems. Hence, query evaluation is carried out as an inference process that aims at computing an aggregated utility of a document by considering its relevance to the query but also the corresponding utility with regard to the user's topics of interest. Experimental results using enhanced TREC collections indicate that our personalized retrieval model is effective.