The Research Of Collaborative Filtering Recommender Algorithms Based On User Interests And Item Features

Information overload is one of the most serious problems caused by information overabundance in the information age. With the fast growing of e-commerce market, e-business enterprise provides plentiful products. The choice of individual is heavily interfered with abundant infomation. To solve the problem of information overload, recommendation algorithms are proposed. Collaborative filtering algorithm is one of the most widely used recommendation algorithms. In real environment, the ratio of item rated by users is very low, which results in a very sparse user-item rating matrix. When the user-item rating matrix is very sparse, some similar users can not be found by collaborative filtering recommendation system and can not be recommended to users.In order to solve the sparse problem caused by user-item rating matrix, two algorithms respectively based on item feature and user interests are proposed, which is:collaborative filtering based on item features (CFBIF) and collaborative filtering based on user interests (CFBUI). CFBIF algorithm improves the collaborative filtering recommendation algorithm by utilizing item’s features. On the e-commerce website, items have many attributes. Items with similar features are usually given similar rate. We design a comprehensive similarity measure that considers two aspects, which are:properties and rates. CFBIF predicts unkown rate of user-item rating matrix based on the comprehensive similarity measure, then fill the matrix with the predicted rate. The proposed algorithm increases the data density of the user-item rating matrix, and alleviates the problem caused by sparse rating matrix. CFBUI improves the collaborative filtering recommendation algorithm by mining the user interests. The items rated by a user reflect the interest pattern of the user. Interest pattern of users vary with time because personal interests usually drift. A time-decay based model for user interests is proposed. CFBUI fills the rating matrix based on the model. This algorithm also alleviates the problem caused by sparse user-item rating matrix. This thesis first discusses the problem caused by sparse user-item rating matrix, and then proposes two algorithms based on item’s features and user’s interests, which fills the sparse user-item rating matrix with predicted value and increases the data density of the user-item rating matrix. Finally, the experiment results show that the propposed collaborative filtering recommendation algorithms are more accurate comparing to traditional algorithms.