Jilei Tian

User interest modeling by labeled LDA with topic features

As well known, the user interest is carried in the users web browsing history that can be mined out. This paper presents an innovative method to extract users interests from his/her web browsing history. We first apply an efficient algorithm to extract useful texts from the web pages in users browsed URL sequence. We then proposed a Labeled Latent Dirichlet Allocation with Topic Feature (LLDA-TF) to mine users interests from the texts. Unlike other works that need a lot of training data to train a model to adopt supervised information, we directly introduce the raw supervised information to the procedure of LLDA-TF. As shown in the experimental results, results given by LLDA-TF fit predefined categories well. Furthermore, LLDA-TF model can name the user interests by category words as well as a keyword list for each category.

Dependency relation based detection of lexicalized user goals

Understanding user goal has played an important role in improving the quality of the search engines. Many previous researches focus on finding prominent statistical features to classify the user goal into Borders taxonomy. But it is difficult to achieve high precision because of the weakness of taxonomy definition. This paper first gives a lexicalized taxonomy for user goal, and then proposes a dependency relation based algorithm to detect lexicalized user goals. To alleviate the sparseness of direct dependency relation, we extend our algorithm to include second order dependency relations. The experimental results show that our extended algorithm can achieve precision of 89% on correctness and 79% on relevance, and thus it outperforms previous related algorithm significantly.