一种改进的缺失数据协同过滤推荐算法

An improved collaborative filtering recommendation algorithm for missing data

协同过滤推荐算法是推荐系统研究的热点,近年来,在亚马逊、淘宝等商业系统中获得应用。在实际应用过程中,协同过滤推荐面临数据稀疏和准确性低的问题。作为推荐基础的用户-产品(项目)矩阵通常非常稀疏(存在大量缺失数据),从而导致推荐结果不准确。文章试图在缺失数据情况下提高协同过滤推荐的准确性,聚焦以下两个方面:(1)用户相似度、产品(项目)相似度计算;(2)缺失数据预测。首先,用增强的皮尔森相关系数算法,通过增加参数,对相似度进行修正,提高用户、产品(项目)相似度计算的准确率。接着,提出一种同时考虑了用户和产品(项目)特征的缺失数据预测算法。算法中,对用户和产品(项目)分别设置相似度阈值,只有当用户或产品(项目)相似度达到阈值时,才进行缺失数据预测。预测过程中,同时使用用户和产品(项目)相似度信息,以提高准确度。在模型基础上,用淘宝移动客户端的数据集进行了验证,实验结果表明所提算法比其他推荐算法要优异,对数据稀疏性的鲁棒性要高。

Collaborative filtering recommendation algorithm has been widely studied, and widely applied in recent years in many business sys- tems, such as Amazon, Taobao, etc. In practice, collaborative filtering recommendation algorithm faces the problem of data sparsity and low accuracy. The user-item matrix, which is the basic of collaborative filtering, is usually very sparse （with a large number of missing data）, and this leads to inaccurate results. This paper attempts to improve the accuracy of collaborative filtering recommendation from two aspects： （ 1 ） the similarity between users and items ; （2） the prediction of missing data. Firstly, we used the enhanced Pearson Correlation Algorithm to improve the accuracy of user, item similarity calculation by increasing parameters. Then we proposed a new method for predicting missing data, which is based on both the information of users and the information of items. In our algorithm, we set similarity threshold respectively for the user and the item, and only when users or items similarity meet or exceed the threshold, the missing data is predicted. In the prediction process, we used both the user and the item similarity information to improve the accuracy of the algorithm. Finally, through the experimental analysis of the data set of Taobao mobile client, we found that our algorithm is superior to other collaborative filtering algorithms, and the robustness of da- ta sparsity is much higher.