Self-Supervised Adaptive Hypergraph Neural Network for Incomplete Multimodal Recommendation

With the explosive growth of multimedia information, multimodal recommendation systems play a crucial role in mitigating information overload. However, a majority of existing models operate under the idealized assumption of complete modal information, largely overlooking the prevalent issue of incomplete modalities. Furthermore, their reliance on static, predefined graph structures and sparse interaction labels limits their robustness and generalization capability. To address these shortcomings, this paper proposes SAHRec, a novel S elf-supervised A daptive H ypergraph Rec ommendation framework. SAHRec introduces a data-driven dynamic learning paradigm that jointly optimizes high-order structures and node representations in an end-to-end fashion. The core of SAHRec consists of two major innovations. First, we design a differentiable hypergraph learner that adaptively constructs optimal high-order topological structures from data, moving beyond the limitations of static or heuristic-based methods. This allows the model to capture more precise and task-relevant global dependencies. Second, we introduce a modality-aware contrastive learning task as a powerful self-supervised signal. By aligning node representations derived from local and global structural views, the model is compelled to learn consistent and robust features even in an incomplete information environment, which significantly enhances its generalization. Extensive experiments conducted on three large-scale public multimodal datasets demonstrate that SAHRec, especially under extreme modality absence of up to 90%, significantly outperforms a wide range of state-of-the-art baseline methods, including strong static hypergraph models. This fully validates the effectiveness and superior robustness of our proposed approach in handling the challenging incomplete multimodal recommendation task. Keywords: Multimodal Recommendation, Incomplete Modalities, Hypergraph Neural Network, Self-supervised Learning DOI: 10.7176/CEIS/17-1-04 Publication date: February 28 th , 2026