Recommendation Algorithm for Federated User Reviews and Item Reviews

Abstract

The recommendation model based on scoring matrix is widely used. Although it has achieved certain recommendation accuracy, it ignores the large amount of semantic information available in the reviews that reflects the user's interests, and the data sparsity problem still exists. In response to the above problems, a two-channel CNN recommendation algorithm (C-DCNN, Combine-Double CNN) that combines user reviews and item reviews is proposed. First, the user and item review texts are vectorized into word vectors, and then the features of users and the items are extracted by using two CNN networks respectively. Finally, the abstract features are mapped to the same feature space through the dot product in the shared layer which aims at predicting the user's rating for a particular item. Experiments on the public datasets of Amazon, Yelp, and Beer show that the C-DCNN model makes full use of reviews to characterize the deep features of users and items. The MSE of the model on different datasets is smaller than other benchmark algorithms. And C-DCNN effectively alleviates the problem of data sparsity.