融合用户和商品评论的双通道CNN推荐算法

Dual channel CNN recommendation algorithm combining user and product reviews

基于评分矩阵的推荐模型目前被广泛应用,虽达到一定推荐精度,但忽略了评论中大量能够反映用户兴趣爱好的语义信息,且数据稀疏性问题依然存在。针对上述问题,提出融合用户评论和商品评论的双通道CNN推荐算法(C DCNN)。首先将用户和商品评论文本矢量化为词向量,再分别使用两个CNN网络对用户和物品进行特征提取,最后在共享层通过点积项将用户和物品的抽象特征映射到同一特征空间,从而预测出用户对特定商品的评分。在Amazon,Yelp,Beer三组公共数据集上进行实验,结果表明该模型在不同数据集上的MSE都比其他基准算法更小,且有效缓解了数据稀疏性问题。

The recommendation model based on scoring matrix is widely used. However,it ignores the large amount of semantic information in the comments that reflects the user′s interests,and the data sparsity problem still exists although it has achieved certain recommendation accuracy. In allusion to the above problems,a Double-channel CNN recommendation algorithm(C-DCNN)that fuses user reviews and product reviews is proposed. The user and product review texts are vectorized as word vectors,and then the users and the items are extracted by using two CNN networks respectively. Finally,the abstract features of the user and the items are mapped to the same feature space through the dot product in the shared layer to predict the user′s scoring for a particular item. The results of some experiments on the public datasets of Amazon,Yelp,and Beer show that the model’s MSE on different datasets is smaller than other benchmark algorithms,which also alleviate the problem of data sparsity effectively.