TextBridgeGNN: Pre-training Graph Neural Network for Cross-Domain Recommendation via Text-Guided Transfer
Researchers have introduced TextBridgeGNN, a novel pre-training and fine-tuning framework designed to enhance cross-domain recommendation systems using Graph Neural Networks (GNNs). Traditional ID-based graph recommendation models struggle with knowledge transfer due to isolated domain-specific ID spaces and structural incompatibilities between heterogeneous interaction graphs. TextBridgeGNN addresses these challenges by utilizing textual information as a semantic bridge to connect disparate domains through multi-level graph propagation. During the pre-training phase, hierarchical GNNs learn both domain-specific and global knowledge, effectively breaking down data islands while preserving collaborative signals. In the fine-tuning stage, a similarity transfer mechanism initializes target domain ID embeddings by leveraging semantically related nodes. Experimental results indicate that TextBridgeGNN outperforms existing methods in cross-domain, multi-domain, and training-free scenarios. Notably, it integrates semantics from Pre-trained Language Models with graph-based collaborative filtering without requiring costly language model fine-tuning or introducing significant real-time inference overhead, offering a more efficient solution for complex recommendation tasks.
Wire timeline
TextBridgeGNN: Pre-training Graph Neural Network for Cross-Domain Recommendation via Text-Guided Transfer
Researchers have introduced TextBridgeGNN, a novel pre-training and fine-tuning framework designed to enhance cross-domain recommendation systems using Graph Neural Networks (GNNs). Traditional ID-based graph recommendation models struggle with knowledge transfer due to isolated domain-specific ID spaces and structural incompatibilities between heterogeneous interaction graphs. TextBridgeGNN addresses these challenges by utilizing textual information as a semantic bridge to connect disparate domains through multi-level graph propagation. During the pre-training phase, hierarchical GNNs learn both domain-specific and global knowledge, effectively breaking down data islands while preserving collaborative signals. In the fine-tuning stage, a similarity transfer mechanism initializes target domain ID embeddings by leveraging semantically related nodes. Experimental results indicate that TextBridgeGNN outperforms existing methods in cross-domain, multi-domain, and training-free scenarios. Notably, it integrates semantics from Pre-trained Language Models with graph-based collaborative filtering without requiring costly language model fine-tuning or introducing significant real-time inference overhead, offering a more efficient solution for complex recommendation tasks.
cs.AI updates on arXiv.org