Text-Attributed Graph Learning

Text-attributed graph learning studies networks whose nodes carry descriptions, documents, or other language, using both these attributes and links for tasks such as node classification, link prediction, anomaly detection, and temporal prediction. Common approaches encode text with pretrained language models and combine it with GNN message passing or graph transformers; newer methods use graph-aware LLM prompts and adapters, contrastive alignment, retrieval, and self-supervised or reinforcement learning. Benchmarks increasingly test few-shot and zero-shot transfer, heterophilic graphs, dynamic networks, and degraded text, edges, or labels, alongside standard citation and product graphs. A central challenge is making semantic and structural signals reinforce each other across domains without letting noisy or misleading inputs harm accuracy, efficiency, or robustness.