基于均值预估的协同过滤推荐算法改进

Improvement of Collaborative Filtering Recommendation Algorithm with Mean Value Estimation

在协同过滤推荐系统中,用户-项目矩阵中存在大量未评分元素,且最终预测值由"最近邻"用户所评分数的加权平均产生。传统算法将未评分元素直接计作0,导致预测得分普遍偏低。针对这种稀疏性引起的问题,提出了一种基于均值预估的协同过滤改进算法。该算法以"最近邻"用户所给平均值对未评分的数据进行估计,有效降低了未评分项目所带来的负面影响。同时该方法又不是单纯的平均值填充,而是在协同过滤算法的第三阶段,需要用到"最近邻"用户对预测项目的评分时,才对"最近邻"评分为0的分值进行替代,这样不会影响到计算的相似度,预测结果不至于平庸。稀疏度为93.7%的数据上的实验表明,在不影响相似度计算的前提下,改进算法可显著降低均方根误差,提高推荐质量;最佳RMSE值可达1.01。

There are a lot of unrated dements in the user-item matrix in collaborative filtering recommendation system,and final prediction value is calculated by the weighted average of nearest-neighbor scores. The values of unrated elements axe regarded as 0 by traditional methods, which results in a generally low prediction score for the unrated elements. In order to solve the problem caused by sparsity, an improved collaborative filtering recommendation algorithm with mean value estimation is proposed. The average score of the nearest- neighbor scores are employed to evaluate the unrated data, effective reduction of negative influence of unrated items in this algorithm. At the same time, this method is not a simple average filling,but in the third stage of collaborative filtering algorithm, when needing to use nearest-ndghbor users to predict, the 0 score is replaced. This won＇ t affect the calculation of similarity, and predicted results not medioc-rity. The experiment results with 93.7% sparse degree show that the recommendation quality can be improved and that the best RMSE value, 1.01, has been reached.