Abstract: The sparsity of data and the lack of negative samples have a bad influence on the result of matrix factorization based one-class collaborative filtering approach, such as features extracted are not obvious and low differentiation between training results. This paper proposes an algorithm to choose positive and negative samples by consi-dering both item similarity and user activity. With this algorithm, more positive samples can be added via the item similarity and more negative samples can be added with the aid of the user activity. In this case, when matrix factorization is used in one-class collaborative filering, the influence of the sparsity and the lack of negative samples can be effectively reduced. A series of experiments show that the new algorithm can add positive and negative samples more accurately, reduce the sparsity, and improve the recommendation accuracy.

Key words: matrix factorization, one-class collaborative filtering, sparsity, positive and negative samples

摘要： 在使用矩阵分解方法解决单类协同过滤问题时，数据的稀疏性以及负样本的缺乏会导致分解特征提取不明确，训练结果区分度低等诸多弊端。针对此问题提出了一种综合考虑物品相似度以及用户活跃度的正负样本选择算法，根据物品相似度向原始数据中添加一定正样本，同时根据用户活跃度向每个用户添加不同数量的负样本，从而减小了稀疏性和缺少负样本对使用矩阵分解方法解决单类协同过滤问题的影响。实验结果表明，该算法能够提高正负样本添加的准确性，减少矩阵稀疏性对单类协同过滤问题的影响，从而提高推荐的准确性。

关键词: 矩阵分解, 单类协同过滤, 稀疏性, 正负样本