融合时间和类型特征加权的矩阵分解推荐算法  

A matrix factorization recommendation algorithm with time and type weight

作　　者：石鸿瑗 孙天昊[1] 李双庆[1] 侯湘[2] SHI Hongyuan;SUN Tianhao;LI Shuangqing;HOU Xiang(College of Computer Science, Chongqing University, Chongqing 400044, P. R. China;Journals Department, Chongqing University, Chongqing 400044, P. R. China)

机构地区：[1]重庆大学计算机学院,重庆400044 [2]重庆大学期刊社,重庆400044

出　　处：《重庆大学学报（自然科学版）》2019年第1期79-87,共9页Journal of Chongqing University

基　　金：国家自然科学基金资助项目(61701051;61472051);重庆市社会科学规划博士项目(2014BS088)~~

摘　　要：针对推荐算法的信息过期问题,结合遗忘函数和信息保持期的改进时间权重引入矩阵分解模型,提出一种基于改进时间权重的矩阵分解协同过滤算法(MFTWCF,MF-based and improved time weighted collabora tive filtering),相比前人提出的基于改进时间权重的邻域协同过滤算法(NTWCF,neighborhood-based and improved time weighted collaboratire filering algorithm),准确性显著提升了26.58%。由于过去的信息所包含的特征在随后的时间里可能被用户持续关注,从而增强过期信息对推荐的影响力,所以提出了融合时间权重和类型影响力加强权重的改进算法(MFTTWCF,MF-bosed and imporved time and type weighteel collaborative filtering)修正上述时间权重。电影数据集的实验证明,MFTTWCF算法预测的准确性比MFTWCF算法提高了3.58%,能够取得更好的推荐效果,适用于通过预测评分进行推荐的系统。In order to solve the problem of information expiration of the recommender systems,we introduced the improved time weight of forgetting function and information retention period into matrix factorization model (MF)and proposed a MF-based and improved-time weighted collaborative filtering algorithm (MFTWCF)whose prediction accuracy had been raised by about 26.58% compared with that of neighborhood-based and improved-time weighted collaborative filtering algorithm(NTWCF). In view of the facts that users could continuously get access to some characteristics of past information, which would have greater influence for recommendation,we proposed the type weight to strengthen the information influence and to correct the improved time weight in MFTWCF. The new improved algorithm is called MF-based improved-time and type weighted collaborative filtering algorithm (MFTTWCF). The results of movie data set experiments show that the prediction accuracy of MFTTWCF algorithm is 3.58% higher than that of MFTWCF algorithm and can achieve better recommendation effect.And it is applicable to recommender systems with rating prediction.

关 键 词：协同过滤 推荐系统 影响力加强权重 信息保持期 时间加权 矩阵分解 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]