Temporal GNNs

Temporal graph neural networks (TGNNs) learn from graphs whose edges, nodes, or attributes arrive or change over time, rather than treating the network as fixed. Common approaches process timestamped events with memory and time encodings, or use snapshot-based recurrent and attention models; newer work adds temporal transformers, adaptive neighbor sampling, motifs, and structural or spectral features. They support future-link and node prediction, anomaly detection, recommendation, and longitudinal applications such as health and finance, often evaluated with chronological splits and dynamic-graph benchmarks. A central challenge is balancing recent changes against long-term memory while remaining scalable and robust to distribution shifts; benchmark scores may not capture semantic drift, learned mechanisms, or whether simple recency and count baselines suffice.