Interpretable Dynamic Graph Learning via Dual‐Granularity Large Language Model Alignment

ABSTRACT Dynamic graph‐level embedding aims to learn whole‐graph representations from temporal snapshots. Existing methods predominantly follow a single‐vector paradigm that encodes each snapshot as a continuous latent vector, suffering from two key limitations: (1) Stage–Snapshot Entanglement, where within‐stage fluctuations and cross‐stage transitions are conflated without explicit granularity distinction; and (2) Implicit Structural Semantics, where meaningful structural properties remain buried in latent coordinates and are never exposed as interpretable, language‐compatible units. To address these issues, we propose DyGLLM, an LLM‐compatible framework that constructs and aligns dual‐granularity structural representations for dynamic graph understanding. Specifically, DyGLLM maps temporally coherent evolution regimes into selective structural tokens , providing explicit stage‐level anchors in the LLM token space. It further constructs graph fingerprints to verbalize each snapshot's prototype deviation, temporal variation and topological profile into language‐readable structural descriptions. A hybrid structure QA alignment mechanism with LoRA fine‐tuning is then introduced to align structural tokens and graph fingerprints with the LLM's representation space, enabling the model to interpret and reason over explicit structural contexts. Extensive experiments on five dynamic graph benchmarks demonstrate that DyGLLM achieves consistently competitive performance across graph similarity ranking, temporal anomaly detection and change point detection, validating the effectiveness of the proposed dual‐granularity structural representation.