RE-DGCL: Reinforcement-evolutionary dynamic graph contrastive learning via proactive decisional evolution

Dynamic graph representation learning is pivotal for modeling evolving complex systems. However, existing methods predominantly follow a "forced evolution" paradigm, in which node states are unconditionally updated upon every interaction. This passive strategy may introduce temporal redundancy and propagate high-frequency stochastic noise, leading to representational drift. Furthermore, an important yet underexplored challenge-termed the "Hard Evolutionary Boundary" problem-is identified in dynamic contrastive learning: conventional random negative sampling may fail to adequately distinguish between recently disconnected historical neighbors and irrelevant noise, resulting in a loose discriminative manifold. To address these limitations, this paper proposes RE-DGCL, a novel framework that reformulates dynamic graph modeling as a sequential decisional control task. RE-DGCL adopts a bifurcated architecture that explicitly decouples Stable Global Embeddings (SGE)-anchoring topological invariants via relational diffusion-from Adaptive Local Embeddings (ALE), which track volatile interaction pulses. To mitigate forced evolution, a Reinforcement Learning (RL)-driven Evolutionary Decision Action (EDA) mechanism is introduced, conceptualizing the node update process as an autonomous information bottleneck: the RL agent learns an "on-demand" gating policy that selectively filters out non-informative fluctuations. Simultaneously, an Evolutionary-aware Negative Sampling (ENS) module is designed to sharpen the contrastive boundary by identifying Hard Negative Constraints (HNC) from a historical proximity buffer. Extensive experiments on eight real-world datasets, together with random-edge perturbation tests, demonstrate that RE-DGCL achieves strong overall performance and exhibits improved robustness in sparse and noisy dynamic scenarios.