AquaGraph and AGN-Net: A Biologically Inspired Framework for Scalable and Adaptive Graph Neural Networks
Abstract
Graph Neural Networks (GNNs) are powerful tools for learning on relational data, but many real-world graphs evolve continuously, and existing Dynamic GNNs (DGNNs) often process large portions of the graph at every time step. In many temporal graphs, change is local: only a small, shifting subset of nodes is active at any step. Motivated by fish-school behavior, where individuals react to local disturbances and nearby neighbors, we propose AquaGraph, a dynamic-graph abstraction, and AGN-Net, a DGNN that implements it through four mechanisms: a predictive-priority signal trained as a self-supervised auxiliary objective, layered aggregation, predictive subgraph retrieval, and localized <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-hop update propagation with a temporal-momentum state update. Our central findings are honestly scoped. The localized-computation mechanisms deliver a <inline-formula> <tex-math notation="LaTeX">$1.84\times $ </tex-math></inline-formula> per-step speedup with no accuracy loss—priority scoring itself remains a full-graph linear pass, so the entire saving comes from restricting message passing to the active region—and the predictive-priority loss improves quality (+1.8 AUC over its removal) by acting as a regularizer, although the learned priority does not anticipate where representation change occurs (Spearman <inline-formula> <tex-math notation="LaTeX">$\approx -0.06$ </tex-math></inline-formula>) and exceeds random selection by only +0.3 AUC. We provide a state-update stability analysis, an information-preservation bound for localized updates, and a worst-case characterization when change is global. On four temporal benchmarks under a faithful, reproducible protocol, AGN-Net is competitive with strong DGNN baselines while using the fewest parameters (56–96K vs. 335–435K for the accuracy leaders); an optional hierarchical-aggregation redesign cuts peak training memory 3–<inline-formula> <tex-math notation="LaTeX">$6\times $ </tex-math></inline-formula>, and precision-weighted (BLUE) aggregation lifts AUC on two of three temporal link datasets. Code and configurations are released for full reproducibility.