DHFI-Net: Dynamic Hypergraph and Fashion Item-Importance-Aware Compatibility Network

The rapid growth of fashion e-commerce platforms has heightened consumer demands for both aesthetic outfit compatibility and the suitability of individual items. However, most research focuses on matching just two items at a time. Some methods use static hypergraphs to model whole outfits, but they rely on fixed rules. These static methods cannot build hypergraphs dynamically. Therefore, they fail to learn hidden compatibility patterns well. Furthermore, the differential contributions of individual items to overall outfit compatibility remain insufficiently explored, limiting the ability to capture the full compatibility effect. To address these problems, this paper proposes a compatibility modeling approach based on dynamic hypergraph construction from a knowledge engineering perspective. The proposed method DHFI-Net frames the outfit composition as a knowledge representation and acquisition problem. Specifically, the hypergraph structure serves as an explicit representation of relational knowledge, where outfits and style groups form hyperedges. The core of our approach, a differentiable hyperedge generator (DHG), acts as a knowledge acquisition process. Building upon a base hypergraph of pre-defined structures, the DHG dynamically learns to discover and refine latent compatibility rules, thereby acquiring vital relational knowledge directly from data. Concurrently, a learnable item importance weighting (IIW) mechanism is introduced to quantify the importance knowledge of each item within a specific outfit context. In DHFI-Net, the intrinsic relationships within outfit pairings are characterized by jointly modeling these two types of learned knowledge. Experimental results on the standard Polyvore dataset demonstrate that the proposed method achieves state-of-the-art performance, with an AUC of 96.03% (disjoint) and 96.17% (non-disjoint) in the compatibility prediction (CP) task, and an accuracy of 77.34% (disjoint) and 78.38% (non-disjoint) in the fill-in-the-blank (FITB) task, outperforming state-of-the-art methods. This work validates not only the efficacy of the model but also provides a generalizable framework for learning high-order relational knowledge in compatibility modeling.