Topic Inspired Model For Recommender System

Nowadays, with the explosive growth of the information, many users face a large number of information which is unrelated to the users, and users are difficult to obtain what they need. To solve this problem, more and more users use recommender system to get useful things. Recommender system is positive about recommending special things to special users and ignoring things which is not useful to special users. As a recommender system, it is the most important thing that the recommending result is accurate, the consuming of the system is low and the time of recommending is fast. These years many researchers try to achieve these goals and they propose many solutions in different ways. Among these solutions the Latent Factor Model(LFM) is the popular topic recently. This thesis present the research and the solution of problem existing in the LFM. The main contributions of the thesis are:1. Summarizing the past research of LFM in the recommender system. This thesis introduces the concept of the recommender system and its common model, analysis the LFM of the recommender system and some improving LFM, and presents the Topic Model that is used in the our model.2. To solve the topic number problem in the Topic Model, this thesis introduces a new method to determine the topic number instead of determining the topic number by experience. And we also prove the result generated by the new method is correct.3. To get the exact meaning of the latent factors in the LFM, this thesis uses Topic Model to get the meaning of the latent factor and the parameter from users’ comment of the items. What’s more, this thesis introduces a Preference Diffusion Method that could revise the parameter in the model to get a more accurate recommendation by the users’ rate on the items.4. We introduce a new Topic Inspired Model with Time, which initializes the items’ parameter matrix with a special value, adds time parameter value to revise the recommendation, and uses the coupling of users’ parameter, items’ parameter and time parameter to update these parameters. The new model’s recommending result is more accurate than existing recommender system with LFM.5. This thesis introduces a new parallel method to solve the problem of consuming too much time during the gradient descent process. And this new method also solves the problem existing in the old parallel method. New method reduces the time of the training process significantly.