Research On Recommendation Algorithm Based On Network Structure

Nowadays, in the academia and industry, there is a word which is frequently talked about, and it is called as big data or mass data. Because of the development of the Internet, various types of massive data about people's work and life are produced. Facing with large amount of data, how to dig out valuable information for people has become a big issue and research subject in the field of information and research. The recommender system that uses collaborative filtering, knowledge mining and machine learning has given users a reliable service and been successfully applied to different web areas of e-commerce, books, movies and videos and so on. The main contents of this thesis are as follows:This thesis briefly studies the similarity model of collaborative filtering algorithm, and the core of collaborative filtering recommendation algorithm is looking for the neighbor users that are similar to the target user. Generally, a simple way of similarity between users is the rating similarity. In fact, it can not accurately reflect the preferences among users that only use rating information, and there exits many other useful information. A main aspect is the feature of item's genres. This thesis combines user's item category preferences with rating information in a weighted way in order to select better neighbor users for target users. The experiments shows that the hybrid method has a good performance on the prediction of target user's ratings.The probability spreading recommendation algorithm based on network structure is a kind of simple and effective recommendation algorithm, which builds user and item binary network according to the relationship between user and item at first, then recommends items to object user according to the analysis on network structure. Nowadays, how to make use of more data features information from the dataset such as item's content, user's personal information and so on is the focus of recommender system in the current research. Scoring information especially can well reflect the user's preference on different items in the perspective of item, and the user's score information can also be used as a reflection of the degree of similarity between item and item. Therefore, based on the item category information and the user's rating respectively, this thesis proposes two kinds of similarity between items in order to combine them with the probability spreading matrix in a proposed way, and finally improves the accuracy of the recommendation algorithm.The item ranking recommendation algorithm based on the network structure is a method based on the restart of the random walk, and the main factor that affects the algorithm is the calculation of the item transfer matrix, which is found after the process of detailed analysis. So the algorithm is improved by adding some adjustable parameters into the algorithm. Finally, the experiments show that the performance of the proposed algorithm can be well improved by choosing suitable parameters.