An Improved Singular Value Decomposition Recommender Algorithm Based on Local Structures
Fang Yao-ning① Guo Yun-fei① Ding Xue-tao② Lan Ju-long①
①(National Digital Switching System Engineering and Technological R&D Center, Zhengzhou 450002, China) ②(School of Software, Tsinghua University, Beijing 100084, China)
摘要 基于奇异值分解(Singular Value Decomposition, SVD)的推荐算法，在预测准确性、稳定性上具有明显优势，但在用随机梯度下降法求解过程中误差下降速度逐渐变慢、迭代次数较多，这极大限制了其在实际项目中的应用。针对这个问题，该文利用评分矩阵的差分矩阵来表征局部结构信息，并作为新的目标函数来优化SVD推荐算法。在MovieLens和Netflix数据集合上的实验结果表明：与经典SVD算法相比，该优化算法能够用更少的迭代次数得到更准确的预测结果；与当前的其他算法相比，该优化算法在预测准确性上仅次于SVD++，在训练时间上具有显著优势。
Abstract：Recommender algorithms based on Singular Value Decomposition (SVD) performs better in prediction accuracy and stability, while the speed of error descent slows down gradually and the iteration number is huge when applying Stochastic Gradient Descent (SGD) to SVD , which greatly limits its usage in actual projects. To address this problem, the difference matrix of the rating matrix is used to represent the local structures and to optimize SVD as a new target function. Experiments on MovieLens and Netflix dataset show that the improved SVD algorithm can obtain better performances with fewer iterations compared to classic SVD. When compared with other state-of-the-art recommender algorithms, the proposed algorithm is only second to SVD++, however within far less training time than SVD++.
2013, Vol. 35 
