基于矩阵分解的协同过滤算法研究

Research on Collaborative Filtering Algorithm Based on Matrix Factorization

针对协同过滤算法中相似度计算方式只考虑单一评分数据从而导致推荐效果不理想、且在数据稀疏条件下推荐结果不全、效率不高等问题,提出一种改进的协同过滤推荐方法,通过矩阵分解,构建用户表征向量来计算用户相似度。首先,构建用户表征矩阵和物品项目表征矩阵。其次,设置损失函数,使用户表征向量与物品表征向量内积拟合评分数据,最后,使用用户表征向量计算的相似度与传统相似度以特定权重相融合。在MovieLens数据集上进行实验,结果表明改进后算法在绝对平均误差MAE上有所提升,在数据稀疏的情况下提高了推荐的准确率。

Aiming at the problems that the similarity calculation method in collaborative filtering algorithm only considers a single score data,resulting in unsatisfactory recommendation effect,incomplete recommendation results and low efficiency under the condition of sparse data,an improved collaborative filtering recommendation method is proposed,which constructs user characterization vector to calculate user similarity through matrix decomposition.Firstly,the user representation matrix and item represen⁃tation matrix are constructed.Secondly,the loss function is set to make the inner product of user characterization vector and item characterization vector fit the scoring data.Finally,the similarity calculated by user characterization vector and traditional similarity are fused with specific weight.Experiments on MovieLens dataset show that the improved algorithm improves the absolute average error MAE and improves the accuracy of recommendation in the case of sparse data.