Self-Supervised Pretraining for Event-Driven Temporal Graphs via Causal Anonymous Walks

Real-world interaction networks form event-driven temporal systems that can be modeled as continuous-time interaction graphs, where learned representations should preserve temporal causality and generalize to evolving or unseen nodes. We propose Causal Anonymous Walk Contrastive Learning (CAW-CL), a self-supervised pretraining framework that constructs a positive pair by independently sampling two sets of causal anonymous walks from the same node at the same reference time. A shared encoder is trained to align the two views using a symmetric InfoNCE-based objective, yielding temporal node representations. After pretraining, the encoder is frozen and reused with lightweight task-specific prediction heads for downstream evaluation. Experiments on Wikipedia and Reddit cover transductive and inductive temporal link prediction and dynamic node classification. Among the evaluated methods, CAW-CL achieves the best link-prediction performance, performs strongly on node classification, and outperforms the architecture-matched Causal Anonymous Walk End-to-End (CAW-E2E) baseline.