IDG: Intrinsic Assessment and Information-Theoretic Learning for Dynamic Graph Neural Networks

Dynamic graph neural networks (DyGNNs) aim to transform evolving graph structures into informative low-dimensional representations to support various downstream tasks, such as link prediction and node classification. However, assessing the quality of latent embeddings remains difficult because existing methods mainly rely on downstream task performance. Such end-to-end assessment lacks interpretable metrics for directly evaluating the latent embeddings themselves. This reliance leads to repeated tuning trials, limited diagnostic insight, and unstable model selection across tasks and datasets. To address this issue, we propose **IDG**, which introduces **I**ntrinsic Assessment and Information-Theoretic Learning for **D**ynamic **G**raph Neural Networks. Specifically, IDG contains two key components: (i) structure mutual information (SMI), which uses spectral entropy to measure structural information shared by inputs and latent embeddings, and (ii) feature mutual information (FMI), which uses $f$-divergence to capture the dependence between latent embeddings and node attribute dynamics. To further demonstrate the practical utility of these metrics, we integrate them into a representation learning objective to guide embedding optimization. Comprehensive experiments show positive empirical associations between the proposed metrics and downstream performance across diverse benchmarks and model architectures, providing interpretable clues about model efficacy, failure modes, and robustness patterns across settings under the evaluated protocols. The code is available at https://anonymous.4open.science/r/IDG_ICLR-924F.