基于多模态用户图的跨领域服务匹配方法OA

In cross-domain service matching scenarios,user preferences are distributed across multiple service domains and exhibit pronounc multimodal and heterogeneous characteristics.Accurately modeling users'cross-domain preferences are therefore crucial for improving matching performance.Existing methods predominantly relied on unified embedding spaces,which tended to compress multi-source preferences and overlooked modality-specific discrepancies.To address these limitations,a multimodal user graph-based cross-domain service matching method was proposed.The proposed ap-proach constructed a multimodal user graph using preference labels as anchors,explicitly modeling user preferences across text,image,audio,and video modalities,while incorporating an open knowledge graph and virtual auxiliary nodes to en-hance graph connectivity.Furthermore,a modality-aware graph pooling module,termed MUGPool,was designed to adap-tively aggregate preferences across different modalities.Experimental results on the Amazon multimodal multi-domain data-set demonstrate that the proposed method outperforms state-of-the-art cross-domain service matching baselines.