Design And Implementation Of Personalized Recommendation Algorithm Of E-business

E-commerce system gives users more and more choices, meanwhile, information overload is grim increasingly and framework of system becomes more complex, then it becomes more and more difficulty for users to find what they like, then e-commerce personalized recommedation system appears.The recommendation algorithm is the core of the recommedation system, and it determines recommendation results to a great extent. Collaborative filtering system gathers ratings from people of the same interest with the target user and then creates recommendations, and it has a high degree of personalization, so it is the most successful and popular method. However, there are still many deficiencies in practical application, such as inaccurate calculation in user similarity, the real-time response, new-item and accuracy problems.This paper proposed a clustering users algorithm based on users preference for item sort to meet the needs of real-time. The algorithm firstly clusters items based on attributes, and gets users preference for item sort. Then it uses k-means clustering to cluster users, and lets the users with the same interest in the same class. We can find the user's nearest neighbor from several nearest clusters to avoid the entire users base,and enhance the real-time response speed.Because the first center of k-means clustering is random, it will result that userclusters are random. This paper uses kruskal algorithm with user difference evaluation on item sort to produce the first centers, and lets the first centers are near to class centers, then gets clusters with high accuracy.The user-based CF algorithm doesn't consider item relevance,which affects the accuracy, and takes the user's interests in different time into equal consideration, which leads to the lack of effectiveness in the given period of time. In order to revolve these issues, this dissertation advances a CF algorithm based on item relevance. The algorithm adds item relevance to calculate user similarity, then avoids disturbance of unrelevant item, at the same time, it adds time as a weight for computing missing ratings, and makes the interests approaching the gathering time have bigger weight in recommendation process.In the end, this paper takes two experiments with MovieLens data sets: experiment of searching for nearest neighbor and CF algorithm experiment. The first experiment uses minimum space searching for more neighbors to estimate result, and experiment results show that k-means clustering based on preference for item sort can find more neighbours from minimal space than k-means clustering, and it improves accuracy of finding neighbour; The second experiment useds MAE to evaluate recommedation quality. Compared the improved recommedation algorithm and traditon recommedation algorithm, experimental results show that the improved algorithm is more precise and gives better prediction in accuracy.