Research On Personalized Service Recommendation Methods In Mobile E-Ecommerce

In recent years, the mobile e-commerce has become the new development trend of the electronic commerce, fuelled by advanced and mature of wireless and mobile technologies and the continuously increasing number of mobile customer.The emergence of mobile e-commerce provides lots of opportunities for the urgent needs of applications in precision marketing, comparison shopping, dynamic supply chain and real time optimization of distribution and other aspects. Meanwhile, it has been known that the mobile e-commerce has several features:mobility, virtualization, sociality, unstructured data and personalization. These features bring series of problems to personalized recommendation technology for mobile e-commerce.Our contributions are shown as follows:(1) The theories and methods of personalized recommendations in mobile e-commerce are overviewed and summarized(2) To overcome the low accuracy and slow convergence of traditional clustering algorithms in customer segmentation, an improved hybrid clustering algorithm named KSP is proposed, which integrates advantages of K-means, SOM and PSO. The initialization of KSP is optimized by K-means and SOM; the solving process is carried out by the combination of PSO and K-means with a mechanism of restraining premature stagnancy. Then, a customer segmentation model was established to analyze types of customers in catering industry under mobile business environment. Also, an actual case was illustrated to verify the efficiency of the KSP algorithm.(3) A novel association rule mining algorithm based on matrix and interestingness (MIAR) to improve both the efficiency of finding frequent itemsets and the quality of rules for personalized recommendations. By constructing a data structure named transaction matrix, MIAR creates candidate itemsets by taking database scan only once and deletes infrequent itemsets to compress searching space to avoid redundant candidate item during mining the frequent itemsets. Meanwhile, MIAR has an improved interestingness measure to prune rule for solving the "rule explosion"problem, which includes both of subjective and objective measures and takes account into the contextual information. In experiments on four different datasets, it is found that MIAR not only outperforms other three algorithms in term of the efficiency, but also enhances the quality of rules by reducing the quantity of irrelevant rules. Thus, MIAR has a better applicability for precise personalized recommendation.(4) With the overload of customers and products information, the ratings matrix has quickly grown into a high-dimensional sparse matrix. Meanwhile, traditional collaborative filtering (CF) methods assume that all users have the same weight on ratings, although the customers'interests and demands are actually context-dependent. These problems severely affected the quality of recommendation. To solve the problems, we propose a novel CF method combining context clustering and social network analysis. Firstly, all users are clustered into different groups by context information, to reduce the sparsity and dimension of ratings data. Then, a user ranking model based on social network analysis is constructed to estimate the correlations between users, and incorporated into similarity measure for improving the quality of recommendation. Experiments results show that the proposed method outperforms other methods and improves recommendation quality effectively.(5) Based on those method in (3),(4) and (5), a personalized recommendation system is designed in this dissertation, and its analysis module is independent of the service module. Therefore, this recommendation system has good portability and maintainability. Meanwhile, those proposed methods in this dissertation are implemented by using the COM technology and embeded in a prototype CRM system system for catering industry under mobile e-commerce environment.This research can boost the study of personalized recommendation in mobile e-commerce to a certain extent, and it provides a theoretical and methodological support for the practical application.