Diffusion Review-Based Recommendation

Review-based recommender systems have attracted a lot of attention recently because of the rich information entailed in the reviews. These models try to learn representations to model user interest and item features from textual reviews. However, most existing methods usually simply learn the mapping relationships between input data and output labels, which could be unable to effectively model the complex user-item interaction generation process and diverse user interests. Inspired by generative models like diffusion models, in this paper, we propose a Review-based Diffusion Recommendation model (RDRec), which aims to learn more real user interest distribution and simulate user-item interaction generation process. In the forward process, RDRec corrupts the review features by adding Gaussian noise and trains the transformer as the approximator to reconstruct the origin features. Besides, the user’s historical behaviors are also fed into the approximator to combine the user’s other interaction information. In the reverse process, RDRec reverses the corrupted user features in a smaller step as the user interest representation. In this way, the model could capture the user’s diverse interests and learn the user interaction generation process. Experiments are conducted on four benchmark datasets and the results validate our model’s effectiveness in recommender systems.