摘要
针对传统的基于用户的协同过滤算法存在的推荐效率、精度和个性化低的问题,提出一种改进方法.该方法在计算用户评分矩阵时,考虑到用户评分矩阵稀疏性,建立项目-用户的倒查表,只计算有相同评分项的用户之间的相似度,避免了传统方法中对所有用户计算两两用户相似度的庞大工作量.该方法在计算用户相似度时,考虑到项目的热门程度不同,"惩罚"了用户共同兴趣列表中的热门项目,避免了传统方法中赋予所有项目相同权值对推荐结果个性化的负面影响.本文在详细分析了改进的用户协同过滤算法的原理和优点,给出了其推荐步骤流程图.在Movielens100K和HetRec2011-movielens-2k公开数据集上,十折交叉验证的结果表明,改进后的算法节约了运行时间,提高了推荐算法的效率和个性化.
Aiming at lowefficiency,precision and personalized of recommendation existing in traditional user-based collaborative filtering algorithm,an improvement recommendation algorithm is put forward in this paper. Considering data sparseness of user-rating-data matrix,this improvement recommendation algorithm introduces items—users inversion table,and only calculates users＇ similarity of whom with the same rating items in user-rating-data matrix. Therefore,it can avoid huge workload,which is brought by the pair-wise users＇ similarity calculation in traditional method. Considering different hot degree of items in users＇ similarity calculation,this improvement recommendation algorithm punishes the influence of popular items on users common interest lists for similarity. Therefore,it can avoid negative influence of popular items in recommendation results personalized,which is brought by giving the same weight of all items in traditional method. After analyzing principles and advantages of improved user-based collaborative filtering algorithm,newrecommendation steps and flowchart are given. The experiment results of 10-fold cross-validation in Movielens100 K and HetRec 2011 open datasets shows that improved algorithm could save running time,increase efficiency and personalization of recommendation.
出处
《小型微型计算机系统》
CSCD
北大核心
2016年第3期428-432,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(51305142
61103170)资助
厦门市科技计划项目(3502Z20143041)资助
福建省自然科学基金计划项目(2014J01191)资助
中国博士后科学基金第55批面上项目(2014M552429)资助
华侨大学引进人才科研启动项目(12BS217)资助
关键词
基于用户的协同过滤
个性化推荐
相似度计算
用户评分矩阵
数据稀疏性
项目-用户倒查表
十折交叉验证
user-based collaborative filtering
personalized recommendation
similarity calculation
user-rating-data matrix
data sparseness
items—users inversion table
10-fold cross-validation
