김태리/박혁주’s paper has been accepted in

Title: Multimedia Recommendation by Integrating Modality-Aware Diffusion and Similarity-Based Graph Refinement,
Author: Taeri Kim*, Hyeokju Park*, Hyunjoon Kim, and Sang-Wook Kim
Abstract
Multimedia recommender systems have recently adopted Diffusion Models (DMs) to generate realistic user–item relationships based on common interaction patterns, thereby improving accuracy. However, we identify two key limitations in existing DM-based multimedia recommenders. (1) When applying the noise injection–removal mechanism of DMs, they inject random noise into all user–item pairs, without considering the multimodal features of users and items or distinguishing between interactions and non-interactions. (2) Although DMs effectively capture common relationships, they fail to adequately capture niche user–item relationships that reflect users’ and items’ fine-grained characteristics—a limitation supported by our preliminary experiments. To address these issues, we propose ASTRO, a novel multimedia recommendation framework, utilizing contrastive learning by integrating modAlity-aware diffuSion-based graph refinement module and modality-aware similariTy-based gRaph refinement mOdule. Extensive experiments on five real-world datasets demonstrate the effectiveness of ASTRO, showing consistently higher accuracy than the best competitors.