摘要： 当前推荐系统多数存在推荐准确性低、受稀疏性影响大且稳定性差的问题，因此，在Coclus聚类算法的基础上，提出一种评分矩阵与联合聚类的推荐算法。通过Coclus联合聚类，利用图模块度最大化理论分别将评分矩阵的行与列分成g类，经过行列变换形成g×g个低秩评分子矩阵，并对低秩评分子矩阵进行矩阵分解，填充缺失值，以提高推荐质量，在矩阵分解阶段采用改进的非负矩阵分解算法，通过引入L1、L2范数分别提高特征值选择能力和防止模型过拟合，并利用坐标轴下降的迭代算法进行参数更新。实验结果表明，与基线算法相比，该算法具有较高的推荐准确率，且稳定性较强。

关键词: 非负矩阵分解, 联合聚类, 推荐系统, 坐标轴下降法, 模块度

Abstract: Most of the current recommendation systems have many defects,such as low recommendation accuracy,being subject to sparsity and poor stability,so we propose a recommendation algorithm based on Coclus joint clustering and non-negative matrix decomposition.Firstly,through Coclus joint clustering,we use the graph modularity maximization to divide row and column of the scoring matrix into g classes respectively,forming g×g low rank scoring submatrices through row and column transformation.Then we perform matrix decomposition on each low rank scoring submatrix and fill in the missing values to improve the recommendation quality.In the matrix decomposition stage,we adopt an improved non-negative matrix decomposition algorithm,introducing L1 and L2 norms respectively to improve the feature value selecting ability and prevent the over-fitting of model.Finally,we use the iterative algorithm of coordinate descent method to update the parameters.Experimental results show that compared with the baseline algorithm,the proposed algorithm has higher recommendation accuracy and better stability.

Key words: non-negative matrix factorization, joint clustering, recommendation systems, coordinate axis descent method, modularity

中图分类号: 

• TP391