基于局部优化奇异值分解和K-means聚类的协同过滤算法  

Collaborative filtering algorithm based on singular value decomposition of local optimization and K-means clustering

作　　者：尹芳[1] 宋垚 李骜[1] Yin Fang;Song Yao;Li Ao(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学计算机科学与技术学院

出　　处：《南京理工大学学报》2019年第6期720-726,共7页Journal of Nanjing University of Science and Technology

基　　金：黑龙江省青年创新人才项目(UNPYSCT-2018203);“理工英才"计划项目(LGYC2018JQ013);黑龙江省自然科学基金(YQ2019F011)

摘　　要：为了克服传统协同过滤(CF)推荐方法数据稀疏和可扩展性差的不足,该文提出1种基于局部优化降维和聚类的协同过滤算法。采用局部优化的奇异值分解(SVD)降维技术和K-均值(K-means)聚类技术对用户-项目评分矩阵中的相似用户进行聚类并降低维度。利用近似差分矩阵表示评分矩阵的局部结构,实现局部优化。局部优化的SVD降维技术可以利用更少的迭代次数缓解CF中数据稀疏和算法可扩展性差的问题。K-means聚类技术可以缩小邻居集查找范围,提高推荐速度。将该文算法与基于Pearson相关系数的协同过滤算法、基于SVD的协同过滤算法、基于K-means聚类的协同过滤算法相比较。在MovieLens数据集上的实验结果表明,该算法的平均绝对误差(MAE)较其他算法降低了大约12%,准确性(Precision)提高了7%。A collaborative filtering(CF)algorithm based on dimensionality reduction of local optimization and clustering is proposed for the data sparsity problem and poor scalability of traditional CF recommendation method.The locally optimized singular value decomposition(SVD)of matrix dimension reduction technique and the K-means clustering technique are used to reduce the dimensions and cluster similar users in a user-item scoring matrix.An approximate difference matrix is used to represent the local structure of the scoring matrix and implement the local optimization.The locally optimized SVD technique can alleviate the problem of data sparsity and poor scalability in CF by using fewer iterations.K-means clustering technique can narrow the search range of neighbor sets and improve the recommendation speed.This algorithm is compared with CF algorithms based on Pearson correlation coefficient,SVD,K-means clustering respectively.Experimental results on the MovieLens dataset show that the mean absolute error(MAE)of this algorithm is about 12%lower than those of other methods,and the precision is 7% higher.

关 键 词：局部优化 奇异值分解 K-均值聚类 协同过滤 近似差分矩阵 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]