CORE-DG: Community role-aware contrastive learning on text-attributed dynamic graphs

Current contrastive learning methods for dynamic graphs derive training signals from topological evolution alone, ignoring the textual content that edges often carry. Incoming edge texts, such as citation contexts, user reviews, and transaction notes, describe how a node is functionally perceived by its community, forming what we call a community-attributed semantic role (CASR). We propose CORE-DG, a contrastive learning framework built around this signal. CORE-DG extracts per-window CASR vectors through a Temporal Attention Role Aggregator, distills a temporally stable role via cross-window attention pooling, fuses it with identity and structural representations through a gated mechanism, and uses the distilled role to guide hard negative mining. Theoretical analysis confirms the gradient advantage of role-guided negatives under mild conditions. On four text-attributed dynamic graph benchmarks, CORE-DG outperforms all thirteen baselines on node classification, improving Accuracy over the strongest prior method by 2.17 points on Cora (86.41%), 3.82 on IMDB (76.27%), 5.20 on Sephora (66.89%), and 4.51 on Dianping (72.35%), with all improvements statistically significant under a paired full-pipeline test ( p < 0.05 ). Ablation and edge-text perturbation studies indicate that the gains originate from the semantic content of edge text rather than additional model capacity.