A Multimodal Heterogeneous GNN-Based Recommendation System for PC Component Selection with Compatibility Constraints

The increasing demand for customizable PC configurations in the e-commerce space presents a dual challenge: personalizing recommendations while ensuring component compatibility. This paper presents a novel recommendation system built upon a Multimodal Heterogeneous Graph Neural Network (MM-HetGNN), integrating diverse data modalities including structured specifications, user preferences, review sentiment, and product imagery. The system encodes users and components as nodes, with edge types representing compatibility, preferences, and multimodal content links. Textual features are embedded using BERT, visual features via ResNet, and graph features are propagated using edge-aware heterogeneous GNNs. Our framework outperforms conventional recommendation systems in both accuracy and compatibility assurance, offering a scalable solution for next-generation PC-building platforms.