Research And Implementation Of International Contents Based Personalized Recommender System

The Internet brings huge data to users. However, it leads to the low efficiency of data usage. Facing the massive web resources, such as videos and news, users cannot acquire the resources they really need. In order to solve this problem, personalized recommender system is proposed. The recommendation is generated by using the history records of users and the recommended web resources help users find what they interest in. Some important problems have not been solved in recommender system. Firstly, there is lacking an effective way to build the relationship between implicit information and user's preference features. The preference is important to the recommeder system, but some system can only get the history behaviors of users, which is also called implicit information. Secondly, though a number of recommendation algorithms are proposed in recent years, most of them can only solve the situation which the explicit information is obtained in the system. Therefore, to solve these two problems, this paper chooses web news resources as the research data, and designs feature modeling algorithm as well as news or text recommendation algorithm.The content in this paper can be summarized as follows: Firstly, a feature modeling algorithm that bridges the gap between the implicit information and user's preference is designed. The encoding method,which is called VLAD (vector of locally aggregated descriptors)encoding, is proposed to model user's preference by using history records. Second, in terms of improving the performance of the encoding method, two modified encoding algorithm are proposed. The experiment results show that both of them can eliminate the influence of initialization and prevent learning from getting into local extreme point.Thirdly, a new recommendation algorithm, namely Siamese Fitting Neural Network, is proposed.The proposed method in this paper has several advantages. Firstly,the method can build preference feature vector through the history records of users. Secondly, the method has great potential to be applied to many systems, since it only needs history records instead of the explicit information. Thirdly, the method has strong flexibility. Though this paper solves the web news recommendation problem, the method can be adapted to any web resources recommendation problem.Therefore, it has potential commercial value.