Meta-Adaptive Continual Learning for Dynamic Social Communication Graphs

Abstract

The rapid and continuous evolution of user interactions in social communication systems such as social media platforms, messaging applications, and collaborative content networks creates highly dynamic graph structures. These evolving environments present a significant challenge for accurately modeling and understanding user behaviors, which is essential for tasks such as personalized recommendations, community detection, and social influence analysis. Traditional graph neural networks (GNNs) are designed for static graphs and struggle in dynamic settings, suffering from catastrophic forgetting as new information overwrites previously learned knowledge. Given the critical need for adaptive models that can capture the complexities of real-time, evolving social graphs, we propose Comm-CLGNN, a novel communication-aware continual learning GNN. Comm-CLGNN introduces a dual-strategy approach: a meta-adaptive task modeling module that enables fast adaptation to evolving graph patterns, and a competitive memory selection mechanism that preserves task-specific knowledge over time without relying on replay buffers or complete retraining. Additionally, we apply a lightweight L2 regularization on task-specific parameters to discourage abrupt changes to useful components, further reducing interference with previously learned representations. We evaluate Comm-CLGNN on multiple real-world benchmarks and show that it achieves consistently stronger average performance and competitive forgetting behavior across datasets. For example, Comm-CLGNN attains $94.89\%$ AP with 0.03 average forgetting on CoraFull, and on Email-EU-core it outperforms strong baselines with backward transfer of $-0.3$ and forward transfer of $+2.0$ .