TextOG: A Recommendation Model for Rating Prediction Based on Heterogeneous Fusion of Review Data

It is beneficial to use user review as a preference expression because they contain information that is not in the interaction record. However, most current research on recommendation systems only models the explicit records of users and items. It does not mine more personalized information from the review texts generated simultaneously with the interaction records. In this article, we proposed a heterogeneous fusion recommendation model for extracting fine-grained product attributes and user behavior from the review texts. The model we proposed is called TextOG. In the first half of the model, we used two blocks to learn user reviews and item reviews, one of which is dedicated to learning user behavior using reviews written by users, and the other block determines product attributes from reviews written for products. In the second half of the model, we connected the latent factors learned by users and items to perform spatial convolution on the graph. That way, implicit features can perform complex interactions in non-Euclidean spaces. We conducted experiments on a large review data set, and the results show that TextOG performs better than the baseline recommendation methods on various datasets.