Research And Application Of Personalized Recommendation Algorithm Based On Conditional Relative Average Entropy

With the popularity of Internet and the rapid development of E-Commerce, online transactions are increasingly popular. It shows that more and more products trading are transformed from real deal into virtual transactions, leading people to be faced with the faster and faster growth of information resources' types and scales. However, it also promotes the research and development of E-Commerce personalized recommendation. C urrently, the E-Commerce personalized recommendation is mainly based on a variety of related relations, such as based on the relationship between the commodities, between the users, between the user and the product. If there are little history data about customers consumption behavior or not, or there are little o verlapping data chosen by customers, or there are not some commodities in the history data, the connections between relative relationship will be short and insufficient, resulting in the data sparsity and cold start problem because of the invalid similarity calculation. Under these circumstances, the accuracy of recommendation is low and it is difficult to provide properly recommendation service to customers. In addition, consumer personality of customers has become gradually prominent. Their preferences and consumption characteristics have an important influence on their consumption behavior. Users will be able to consume when the utility of the products are in line with the consuming characters. This offers a new research vision to e-commerce personalized recommendation.Therefore, how to accurately grasp the users' interests, how to provide customers high quality and high efficient recommendation service and comfortable experience, become the problems to be solved by the site.The mainly work of this paper includes:(1) A deta il analysis of the personalized recommendation algorithms and the complex network community structure discovery algorithms as well as characteristics of each algorithm.(2) Taking into account the demand of accuracy about personalized recommendation system nowadays, the thesis chooses CNM algorithm that is representative in recommendation algorithms, and optimize CNM algorithm through the link weight, the vertex weight and JSD distance formula, and then confirm the improved algorithm.(3) With the analysis of the characters of customers, and bringing with conditional mutual information and conditional relative average entropy, the new algorithm is used to obtain the initial input order of nodes of K2 algorithm, and then study the Bayes ian Network from CH score function and posterior probability function, so as to analysis the consuming characters.(4) Infer and judge the recommended list matched with consuming characters by Junction Tree algorithm on the existing Bayesian Network, in order to get the final recommendation domain.(5) A detailed description of Telecom Leisure Assets Market-Trade System, including requirement analysis and system design. Finally, the research production is applied to assets recommendation modules.