A fabric graph construction method based on subdivided attribute decoupling and hierarchical modeling for fabric recommendation

To address the contextually rooted industrial pain points in the textile industry, including fragmented fabric attribute representation, inconsistent definitions of fine-grained attributes, difficulty in structurally describing attribute relationships, and low efficiency in knowledge retrieval and reuse in the textile industry, this study proposes a fabric graph construction method based on subdivided attribute decoupling and hierarchical modeling for fabric recommendation. The approach systematically organizes the multilevel fiber-to-fabric attribute system and standardizes fine-grained information, including fabric composition, structural parameters, weaving processes, and functional applications. On this basis, it integrates heterogeneous sources, including enterprise standards, industry literature, and product data, to design a fiber-oriented ontology and relational schema, thereby constructing a knowledge graph with detailed attributes. Finally, a recommendation model is developed based on knowledge paths, incorporating contrastive learning and self-attention mechanisms to optimize knowledge path representation and enable dynamic semantic modeling, supporting relational reasoning and knowledge application. Experimental validation on enterprise fabric datasets (Sun fabric dataset, SFD) and the Amazon fabric dataset (AFD) demonstrates that the proposed method achieves improvements, reaching a hit ratio (HR) @50 = 0.9056 and normalized discounted cumulative gain (NDCG) @50 = 0.8675 on SFD and HR@50 = 0.7996 and NDCG@50 = 0.5902 on AFD, with consistent improvements over all baseline models in real textile application scenarios.