| Recommender systems have been widely used, and it is important to make them more perfectly. This thesis made a research centering on the relevant problems:the description of the content of the products, the computation of Ratings, the change of user’interest, and the diversity of recommendation results etc. The main contribution can be concluded as follows:(1) The image features are used in the recommender system, which solved the problem of ambiguity existing in the way of presenting the product’s appearance using text words. The method of "bag of words" is proposed to describe the product, which can describe the product completely both the attribute characteristics and the image features in a unified form.(2)The GC-PLSA (General Content Probabilistic Latent Semantic Analysis) model is proposed to solve the new product problem existing in other techniques of collaborative filtering. Two latent variables representing the user groups and product groups respectively are contained by the GC-PLSA model, which is trained by using the asymmetric learning method; GC-PLSA can improve the accuracy of the predicted ratings and solve the new product problem effectively.(3)The concept of "User Interest Factor" is proposed, which can deal with the change of user’s interest for a certain type of product before or after he (or she) buys it. By introducing this factor to the ratings, the recommendation results can be more accurate.(4)A new re-ranking model called ITRM (Improved Traditional Ranking Model) is proposed to improve the aggregate diversity for all users in the system. ITRM model can improve the aggregate diversity effectively with only a small amount of accuracy loss, and it is parameterized, namely the relationship between the diversity and the accuracy can be controlled by regulating the parameter of the model.The experimental results showed that the methods proposed in this thesis can improve the recommendation quality of the product recommender system effectively. |
PDF Full Text Request