基于降噪自编码和卷积神经网络的协同过滤算法

Collaborative Filtering Algorithm Based on Denoising Auto-encoder and Convolutional Neural Networks

在协同过滤算法中最主要是用户相似度计算,但是用户评分项数据存在严重稀疏,导致推荐精准度降低。针对评分项数据稀疏性问题,论文提出一个C-DAE协同过滤算法。首先,利用卷积神经网络(CNN)对项目评论文本提取用户兴趣偏好,得到项目向量矩阵,其次,利用项目向量矩阵对降噪自编码器(DAE)加权填充原始评分矩阵,最后填充后的评分矩阵计算用户相似度进行推荐。实验结果证明,该方法解决了评分项数据稀疏性问题,提高了推荐质量。

The most important step in the collaborative filtering algorithm is the user similarity calculation.But the data of user rating items is severely sparse,and result in lower recommendation accuracy.This paper proposes a C-DAE collaborative filtering algorithm to solve the problem of sparsity rating data.Firstly,by using a convolutional neural network(CNN),user preferences for item review texts can be extracted,and a item vector matrix can be obtained.Secondly,the item vector matrix is used to be the initial weights of the denosing auto encoder(DAE)to fill the original rating matrix.Thirdly,the filled rating matrix id used to calculate the user similarity for recommendation.The experiment results show that the proposed method solves the sparsity problem of the rating data and improves the recommendation quality.