Abstract: The inherent data sparsity and cold start problems in probabilistic matrix factorization(PMF) limit the effect of academic paper recommender. To remedy the shortcomings and enhance the recommender effect, a new hybrid recommender model named as matrix factorization with topic(MFWT) was proposed. The model constructs topic characteristics of both users and papers using the author-conference-topic over time(ACTOT) model and the traditional latent dirichlet allocation topic model respectively, enhancing the corresponding user and paper latent factor characteristic vectors of PMF model. Experiments show that the model well overcomes the data sparsity problem and the cold start problem of PMF and increases the effect of academic paper recommender.

Key words: probabilistic matrix factorization, topic model, hybrid recommender model, data sparsity