Research On Users' Clustering And Users' Interest Changing Issue In E-commerce Recommendation Systems

Among all recommendation systems, Collaborative filtering recommendation technology is the most widely used one. However, as the scale gradually expands, recommendation system is now facing a great challenge on both real-time responsiveness and recommendation performance. In order to addressing the real-time responsiveness issue, Collaborative filtering technology, based on clustered users using K-Means Algorithm divided recommendation process into two parts: the offline part and the online part. In offline part, K-Means Algorithm partition all users into several clusters, while in online part, systems make recommendation according to cluster which the certain user belonged to. This approach would improve real-time responsiveness for recommendation system, but unfortunately has its own flaws: on the one hand, the algorithm requires an initial partition and the condition of initial partition has close relationship with the quality of clustering result; on the other hand, the clustering result exist local optimum issue. These defects will definitely affect the performance of the algorithm. Thus, in order to overcoming these problems, this paper proposes using AntClass clustering algorithm instead of K-Means Algorithm. AntClass clustering algorithm does not require any initial information and complicated parameter setting. Avoiding the complexity of the algorithm itself, AntClass clustering algorithm can apply to the actual situation much better,and the clustering results are more reasonable.Another problem comes with expand scale to the system is the user's changing interests over time. If the systems use all historical rating data as the traditional collaborative filtering technology did, the quality of recommendation will decrease, since scores provided by user long time ago will have less predictive value. In order to addressing such issue, this paper proposes taking all rating data as data stream, and then store this data stream with the technology of pyramid time frame. As a result, the farther away from the current time, the lower availability the data will have, and in another word, the closer from the current time, the higher availability the data will have.In a word, this paper store data stream with the technology of pyramid time frame, and cluster all users with AntClass algorithm, so we name such technology as AntStream clustering Algorithm. The final experimental results indicate that AntStream clustering Algorithm can not only guarantee the real-time responsiveness, but also improve the recommendation performance to a great extent.