基于聚类的优化协作过滤技术

Research on clustering-based collaborative filtering

基于内存的协作过滤算法主要利用用户对某站点项目的评分,计算2个用户之间的相似性,但该方法可扩展性差.基于模型的协作过滤算法通过训练数据预先计算出预测模型,弥补了上述方法的不足,但该模型没有考虑到个体的差异而限制了推荐的性能.在总结现有2种算法特点的基础上,提出一种新颖的协作过滤框架,它先从训练集中产生聚类,并以此为基础进行邻居预选择,再在预选择的邻居集合上使用基于内存的协作过滤算法.实验结果表明,该方法不仅提高了计算的效率,而且也提高了推荐的质量.

Algorithms about memory-based collaborative filtering mainly use the rating of the user to the items to compute the similarity between two users. However, such algorithms are deficient in scalability. Algorithms about model-based collaborative filtering alleviate it through pre-training the model. But the algorithms ignore the diversity of different users. This paper proposes a algorithm to combine the advantages of two algorithms. The users are clustered in advance and the neighbors are pre-selected. Then, the memory-based collaborative on a subset of the users are done. Experiments show that the proposed approach not only improves the efficiency of the computation, but also improves the quality of the recommendation.