A Distributed Robust Out-of-Distribution Consumer Recommendation System Using Diffusion Model

Jiayu Bao, Hongjian Shi, Ziang Zhou, Shiquan Wang, Ruhui Ma, Yang Yue, Zhiwei Song, Haibing Guan, Yuan Liu

Abstract: With the continuous development and widespread application of consumer electronic products, intelligent generation in the field of consumer electronics is capable of generating content based on user preferences and behaviors, providing a more thoughtful and boundary-pushing user experience. Diffusion models, owing to their powerful data distribution modeling capabilities and high-quality project generation abilities, have emerged as a potential study avenue in the domain of recommendation systems. Currently, graph-based recommendation methods using distributionally robust optimization (DRO) assign greater weight to the noise distribution during training, which leads to model parameter learning being dominated by noise. When the model overemphasizes fitting noisy samples in the training data, it may learn irrelevant or meaningless features that do not generalize to out-of-distribution (OOD) data. We propose a diffusion-based distributed robust graph model (DiffDRG) to tackle this issue for ood recommendations. Our solution initially employs a straightforward and efficient diffusion paradigm to alleviate noise effects in the latent space. Additionally, we introduce an entropy regularization term in the DRO objective function to avoid the appearance of extreme sample weights in the worst-case distribution. To assess the efficacy of our system, we perform comprehensive experiments on three datasets across two common distribution shift scenarios.

External IDs:doi:10.1109/tce.2025.3624922