基于Bregman联合聚类与加权矩阵分解的融合推荐算法

A Fusion Recommendation Algorithm Based on Bregman Co-clustering and Weighted Matrix Approximation

针对当前大数据背景下推荐系统中所存在推荐效率低下、扩展性差、推荐质量不高等问题,提出一种基于Bregman联合聚类与加权矩阵分解的融合推荐算法(CO-CWMA)。首先,通过Bregman联合聚类挖掘出多样、不同层次的低秩评分子矩阵,组合不同约束与距离的聚类结果训练得到子模型,进而在各个模型的子矩阵上并发地进行矩阵分解,最后将各个子模型进行均值融合,提高推荐质量、效率与扩展性。在矩阵分解阶段采用SVD++算法,基于每个子矩阵中的评分分布计算加权策略,给予高频评分较大权值,在梯度下降阶段利用学习率函数控制学习率的更新。实验结果表明,与三种基线算法相比,该算法在均方根误差(RMSE)与平均绝对误差(MAE)上均有明显降低,即推荐质量有较大提升。

Aiming at the problems of low recommendation efficiency,recommendation quality and poor expansibility in the recommendation system,a fusion recommendation algorithm based on Bregman co-clustering and weighted matrix approximation(CO-CWMA)was proposed.Firstly,the Bregman co-clustering was used to mine the low-rank sub-matrices of different levels,clustering results of different constraints and distances were combined to train the sub-models,then the matrix approximation was performed concurrently on the sub-matrices of each model.Finally,each sub-model performed mean fusion to improve recommendation quality,efficiency and scalability.SVD++algorithm was used in the matrix approximation stage,weighting strategy was calculated based on the score distribution,and high-frequency score got a larger weight.Learning rate function was used to control the learning rate during the gradient descent phase.Experimental results show that the proposed algorithm has a significant reduction in root mean square error(RMSE)and mean absolute error(MAE)compared with the three baseline algorithms,that is,recommended quality is greatly improved.