An Explainable Recommendation Method based on Diffusion Model
Yupu Guo, Fei Cai, Honghui Chen, Chonghao Chen, Xin Zhang, Menxi Zhang · 2023
To meet the requirements of precise recommendation in the era of information warfare, there has been a growing interest in explainable recommendations in the field of command and control in recent years. While Variational Autoencoders (VAE) have been effective in generating explanations, diffusion models have shown superior performance in fields like image processing and natural language processing. This paper proposes a Diffusion-based explainable recommendation model (DIEXRS) to enhance the effectiveness of explanations further. The DIEXRS model consists of two processes: diffusion (training) and reverse (inference). User and item embeddings are generated based on their IDs, along with comment embeddings from user comments. These embeddings are incorporated into the diffusion process, where the Transformer model is trained to fit the noise and improve the decoder’s decoding capabilities. In the reverse process, a random Gaussian sample and the user’s ID are used to obtain the item’s ID and corresponding interpretation embedding, providing the recommended interpretation through the decoder. Experimental results on two public datasets demonstrate that the DIEXRS model achieves state-of-the-art performance in interpretable recommendations compared to other baselines. This highlights the potential of diffusion models in enhancing the explainability and effectiveness of recommendation systems