Personalizing Web Search Results Based on Subspace Projection

Personalized search has recently attracted increasing attention. This paper focuses on utilizing click-through data to personalize the web search results, from a novel perspective based on subspace projection. Specifically, we represent a user profile as a vector subspace spanned by a basis generated from a word-correlation matrix, which is able to capture the dependencies between words in the “satisfied click” (SAT Click) documents. A personalized score for each document in the original result list returned by a search engine is computed by projecting the document (represented as a vector or another word-correlation subspace) onto the user profile subspace. The personalized scores are then used to re-rank the documents through the Borda’ ranking fusion method. Empirical evaluation is carried out on a real user log data set collected from a prominent search engine (Bing). Experimental results demonstrate the effectiveness of our methods, especially for the queries with high click entropy.