Research On Personalized Recommendation Systems In E-Commerce Based On Data Mining Technoiogy

With the fast development of internet and computer science technology, the model of economic transactions has changed from traditional business type to electronic commerce type. Electronic commerce, which brings much convenience and efficiency to people, has simplified traditional economic transactions in a great extent. However, as time goes by, more and more customers and items are accumulated on the E-commerce system. At the same time, transaction information, access logs, and other business data increases rapidly. As the increase of items, customers of E-commerce system should spend more time to find a product, but the system can only waits for people’s access passively. To solve this problem, some people proposed personalized recommendation system, which can take the initiative to recommend products or information according to users’interest.Currently, there are many recommendation algorithms used for personalized recommendation system, but the most widely used algorithm is collaborative filtering (CF). Compared with other recommendation methods, CF has some advantages, for example, it doesn’t need professional knowledge, and it can also handle unstructured items. On the other side, CF is confronted with some problems, such as scalability, sparsity, cold start and so on. That is the limitation of a single algorithm, neither CF algorithm nor content-based algorithm has its’own limitation, it can give out recommender list but the result is not very precision. To compensate for the defect of a single algorithm, some recommendation systems combine one algorithm with another, even several algorithms combined together. Such as CF method is combined with content-based method, data mining techniques is combined with CF method. Data mining is a good approach to find useful and exact information from a huge amount of data; it provides the basis for business decision. Now many data mining techniques are used in E-commerce Recommendation System, the representative ones are cluster analysis, correlation rules, artificial neural networks and genetic algorithm. In this article’s improvement part, we used Radial Basis Function Network (RBFN) which is one kind of artificial neural networks.The innovation of this paper is proposing an intelligent integrated improvement program, this program includes two parts:algorithm design and model design. The algorithm design is based on the traditional CF recommendation algorithm, using RBFN technique to smooth the sparsity rating matrix, after this smoothing process, the original sparsity matrix has changed to a dense matrix; and then clustering users. The model design is based on the algorithm design part’s clustering result, finding the closest positive and negative neighbors of the active user, and then predicting the ratings of the unrated items which are leading to user’s like and unlike items sets; finally minus the two sets as the result to be recommended. At the end of this article, an experiment is set to test the proposed program’s effect on the book-crossing set by the measure method of MAE and F-Measure. The result shows that the recommendation quality of the improved model is better than the traditional CF recommendation system and the recommendation system based on SVD. This improved program has solved the scalability, sparsity, cold start problems in a certain extent.