SIHG-Rec: Unleashing the Power of Semantic and Interactive Homogeneous Graphs via Dual-Stage Fusion for Multimodal Recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recent studies in multimodal recommendation, which leverage diverse modal information to address data sparsity and enhance recommendation accuracy, have garnered significant interest. Two critical processes in this domain are modality fusion and representation learning. In representation learning, existing studies often extend beyond traditional heterogeneous graphs by constructing homogeneous graphs that incorporate multimodal information, aiming to capture finer-grained user preferences and item properties. However, these studies predominantly rely on semantic homogeneous graphs, overlooking interactive signals, which limits their capacity to capture high-quality user-item relationships. Moreover, prevailing modality fusion strategies are typically single-stage, employing either early or late fusion with simplistic attention or predefined strategies. Early fusion often impedes the learning of modality-specific features, while late fusion fails to sufficiently capture inter-modal relationships. In previous studies, modality fusion and representation learning have been treated as independent processes. In this paper, we argue that these two processes are complementary and can mutually enhance each other: powerful representation learning strengthens modality fusion, while effective fusion improves representation quality. To this end, we propose SIHG-Rec, a novel framework that jointly incorporates semantic information and historical interactions to construct S emantic and I nteractive H omogeneous G raphs, leading to more accurate and robust representations of user preferences and item properties. To fully exploit these graphs, we introduce a dual-stage fusion strategy that first refines all modalities using behavior-guided calibration at an early stage, then fuses their representations at a later stage. This design preserves modality-specific features while reducing task-irrelevant information through early-stage calibration. Furthermore, we adopt an adaptive optimization to ensure balanced and stable representations across modalities. Extensive experiments on three widely used datasets show that SIHG-Rec consistently outperforms a variety of strong baselines, especially in cold-start settings. Additional analyses confirm its superiority under varying levels of data sparsity.