基于SVD与SDAE的神经协同过滤算法

Neural collaborative filtering algorithm based on SVD and SDAE

本文提出一种结合奇异值分解SVD和堆栈式降噪自动编码器SDAE的神经协同过滤算法(NSSCF),利用神经网络有效的高阶特征表示学习能力来提高推荐的质量。NSSCF算法首先通过SVD将原始用户-项目评分矩阵降维,融入辅助信息再用SDAE获取项目特征并计算基于评分的项目间相似度;在项目属性矩阵上计算基于属性的项目间相似度,求出项目间的综合相似度;最后获取待评分项目的最近邻集合并进行推荐。在真实数据集上,经过广泛的实验验证,本文提出的NSSCF算法在很大程度上克服了数据稀疏性问题,在性能上优于其他的传统推荐算法。

This paper proposes a combination of singular value decomposition(SVD) and stacked denoising autoencoder (SDAE). The neural collaborative filtering based on SVD with SDAE (NSSCF) uses the high-order features of the neural network to express learning ability to improve the quality of recommendation. The NSSCF algorithm first reduces the original useritem rating matrix by SVD, integrates the auxiliary information and then uses SDAE to obtain the item features and calculates the similarity between the items based on the rating;then the attribute-based inter-item similarity on the item attribute matrix and the comprehensive similarity between items were calculated;finally, the nearest neighbor set of the item to be scored is obtained and recommended. After extensive experimental verification, the NSSCF algorithm proposed in this paper overcomes the data sparsity problem to a large extent and is superior to other traditional recommendation algorithms in performance.