Bridging Topic Modeling and Personalized Search

This work presents a study to bridge topic modeling and personalized search. A probabilistic topic model is used to extract topics from user search history. These topics can be seen as a roughly summary of user preferences and further treated as feedback within the KL-Divergence retrieval model to estimate a more accurate query model. The topics more relevant to current query contribute more in updating the query model which helps to distinguish between relevant and irrelevant parts and filter out noise in user search history. We designed task oriented user study and the results show that: (1) The extracted topics can be used to cluster queries according to topics. (2) The proposed approach improves ranking quality consistently for queries matching user past interests and is robust for queries not matching past interests.