摘要
针对在协同过滤算法中,传统矩阵分解技术在降维过程中会破坏数据相邻结构的问题,提出基于结构投影非负矩阵分解的协同过滤算法(CF-SPNMF).该算法包含离线学习和在线搜索2个阶段.在离线学习阶段,通过对用户评分矩阵的投影非负矩阵分解,同时保留用户特征的聚类结构,得到低维的用户潜在兴趣因子.在线搜索阶段,将用户潜在兴趣因子进行余弦相似性匹配,发现目标用户与训练样本用户之间兴趣最相似的邻域集合.在实际数据集上的实验结果表明,提出的CF-SPNMF算法与单纯使用矩阵分解和单纯在原评分矩阵上进行用户聚类的推荐算法相比,能够更有效地预测用户实际评分.
In collaborative filtering algorithm,the classical matrix factorization may destroy the adjacent structures among data points from high dimension to low dimension.A novel collaborative filtering algorithm based on structured projective nonnegative matrix factorization（CF-SPNMF）was proposed in order to overcome the problem.The algorithm contains both offline learning and online searching.In the offline learning stage,projective nonnegative matrix factorization was applied to obtain the low dimensional latent factors of user preference without changing the intrinsic structure of users＇cluster.In the online searching stage,cosine similarity was used to measure the similarity between the target user and training users based on the latent factors inferred in the offline stage.Then the most similar neighbor set was further found.The extensive experiments on real-world data set demonstrate that the proposed CF-SPNMF achieves better rating prediction performance than traditional methods using either matrix factorization or users clustering in original rating matrix.
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2015年第7期1319-1325,共7页
Journal of Zhejiang University：Engineering Science
基金
浙江省自然科学基金资助项目(Y1101359)
国家科技支撑计划资助项目(2011BAD24B03)
关键词
协同过滤
投影非负矩阵分解
相邻结构
聚类
collaborative filtering
projective nonnegative matrix factorization
adjacent structure
clustering
