Research Of Movie Recommendation Method Based On Multiple Strategies

In the age of information explosion, the amount of resources provided by the Internet for users is increasing dramatically. The network provides users with a variety of resources as well as difficulties to get access to the information they are really interested in. Personalized recommendation system aims to recommend items to users. It can reduce users’querying and filtering operation, and improve the utilization rate of resources while enhancing user experience. So personalized recommendation has become a hot issue of Internet application research.Personalized recommendation algorithm is the core of personalized recommendation system. It utilizes the user’s past behavior records and characteristics of items themselves to predict the user’s interest in items. Current research and application mostly focus on collaborative filtering and content-based method. To solve the shortcomings such as sparsity, cold-star and low accuracy problem in existing methods of personalized recommendation, this dissertation carries out research on multiple strategies. The main work and achievement of this dissertation is as follows:1. Propose a movie recommendation method based on website relevant recommendation aggregation and movie ontology knowledge. Firstly, a movie ontology model is set up on the basis of movies and celebrities’properties. The user interest model is built by user’s behavior records. And a method is proposed to acquire the weights of user preference of different properties. Secondly, a movie set which is relevant to the movies that the user have watched recently is get to be recommended from some websites by aggregating technology. Thirdly, the similarity between movies or between movies and user model is measured using SimRank method and the weighted average for recommendation. Experimental results show that the accuracy of this method is improved by about10%comparing to the existing methods when it is used to show a recommendation list to user. And the quality of relevant movie recommendation has been improved significantly by11.4%. In some extent, sparsity and cold-start problem can be solved.2. Propose a rating prediction method based on bipartite network and user clustering. At first bipartite network is utilized to represent user’s behavior records of movie items. Use structural features of bipartite network to measure the similarity of users. Then time parameter and the difference between ratings are used to correct the similarity between users in order to acquire the correlation between users. Users are clustered into interest groups based on spectral clustering. At last user’s rating on unknown movies are predicted by neighbor users’ information in interest group. The validity of the method is verified on the standard data set. The experimental results show that the mean absolute error of proposed method is lower than the contrast methods by10%.3. Apply proposed methods in the recommendation module of HappyTV system. Recommendation forms consist of user’s personalized recommendation list and on-demand video’s relevant recommendation list.