Research And Implementation Of Recommendation System Based On Heat Conduction And LDA

With the rapid development of the Internet,big data gradually penetrates into all walks of life. How to retrieve useful information from vast amounts of data has always been an important issue for researchers. The initial information retrieval technology, asking people to provide a keyword, and then return relevant information, but did not meet people for individual needs of the results. In this context, recommendation system came into being, and it can use the preferences of potential users to recommend related information.Recommended system from the beginning to now, has been produced a large number of related applications and research. There are many algorithms such as content-based, association rules, collaborative filtering-based recommendations, etc., but with further development, these algorithms cannot meet the needs of personalized recommendation, because most of the algorithm does not consider context information. Existing recommendation algorithm which based on the user’s past preferences and calculating similarity between users or objects, often meets data sparse or cold start problems, and diversity is not guaranteed. To this end, we study the latest heat diffusion algorithm, and combine with the user’s interest in its improvements. Then we model the users’ interests, and propose a method to extended interests. Finally we propose a hybrid strategy to get better accuracy.First, the paper analyzes the nature of bipartite graph, and decisions that affect user behavior factors were studied. On the basis of consideration of the user’s "active" and users’ interests, heat diffusion algorithm is improved, and the beta test on the two datasets shows that the improved algorithm can obtain higher accuracy.Secondly, combining probabilistic topic model, we model users’ interests, but the presence of the volatility of the interest, the interest of users with existing algorithms to get, can only get the user’s current interest distributions, or get people with similar interests, which cannot predict and mining users’ interests in new. The results are confined to the user known or similar results, which is known as system "over-fitting" phenomenon. To solve that problem, we propose an extension of interest-based random walk algorithm, and experiments show that the algorithm interests expanded can grasp users’ current interests.Finally, we propose two linear mixing algorithms and artificial neural network based on hybrid algorithm to predict the user’s score.