A Nonparametric Mixture Model for Personalizing Web Search

Abstract

Probabilistic topic models were successfully used to achieve the personalization task using query logs. Thus, both users and previously clicked results are considered when estimating probability distrubutions in order to answer users’queries. However, the proposed models are generally parametric and require to define in advance the number of topics. Moreover, they can not deal with new users. To overcome these limitations, we propose a model called the Hierarchical personalized Dirichlet Processes (HpDP) that personalizes search and allows to automatically learn the number of latent topics. It also addresses the challenging problem of predicting results for new users. We compare our model, with recent topic models and use them to rank online products by their likelihood given a particular user/query pair. Experiments performed on data from a real online products comparator show the effectiveness of our approach.