基于稀疏数据预处理的协同过滤推荐算法

摘要

随着推荐系统规模的不断扩大，用户-项目评分矩阵呈现出极端稀疏性，导致基于传统相似性度量方法的协同过滤推荐系统推荐质量的下降。针对该问题，文中提出了一种基于项目特征属性的稀疏数据集预处理方法来提高算法的推荐质量。首先，通过引入项目的特征属性信息，根据项目间特征属性相似度，初步预测用户对未评分项目的评分，可以使得用户-项目评分矩阵完全饱和。接着再对稀疏数据集的未评分项目进行混合填充预处理，避免了传统均值填充法中的用户对项目的评分不可能完全相同的问题以及众数填充法中的“多众数”和“无众数”问题。实验结果表明，文中提出的方法更能有效地提高推荐系统的推荐质量和推荐覆盖率。%With the continuous expansion of recommender systems,the sparsity of the user-item matrix can deteriorate the performance of the traditional similarity calculation based collaborative filtering recommendation approaches. In order to overcome this drawback,a new sparse data pre-processing algorithm based on item feature is proposed to mitigate this effect. First,considering the item characteristics in-formation,the ratings of the unrated items are predicted through the similarities between each item. It can lead to saturated matrix and o-vercome the drawback of the sparsity matrix. Next,the hybrid filling method is utilized to process the unrated items in the sparse data sets,which can avoid the problem of full no consistency of different items for traditional mean-filling method and the multiple mode and no mode for the mode-filling approaches. The simulation demonstrates that the proposed algorithm can improve the recommended quality and coverage dramatically.