Abstract: This paper focuses on the issues that the impact of all users’ historical feedback information has been taken into account when calculating the similarities between any two items, and the traditional collaborative filtering algorithms only utilize the user’s rating data to calculate the similarities. To solve the above problems, two novel algorithms are proposed. Top-N Collaborative Filtering recommendation algorithm based on user Spectral Clustering（SC-CF）：it clusters all uses into several partitions using spectral clustering, then recommends for users in each of the clusters. SC-CF+, compared to SC-CF algorithm, draws the time difference weight, users’ common rating weight and popular items weight into the similarities formula respectively. Experimental results show that the two new algorithms improve the recall of recommendation results.

Key words: collaborative filtering, similarity, historical feedback information, spectral clustering, recall

摘要： 传统的协同过滤推荐算法为目标用户推荐时，考虑了所有用户的历史反馈信息对物品相似度的影响，同时相似度的度量仅依靠用户评分信息矩阵，导致了推荐效果不佳。为解决上述问题，提出了基于用户谱聚类的Top-N协同过滤推荐算法（SC-CF），即应用谱聚类将兴趣相似的用户分成一类，具有相似兴趣爱好的用户比其他用户具有更高的推荐参考价值，然后在类中为目标用户推荐。SC-CF+算法在SC-CF算法的基础上，在相似度度量方法中分别引入了物品时间差因素、用户共同评分权重、流行物品权重。实验结果表明，提出的两种算法提高了推荐结果的召回率。

关键词: 协同过滤, 相似度, 历史反馈信息, 谱聚类, 召回率