TAAM:Inductive Graph-Class Incremental Learning with Task-Aware Adaptive Modulation

TL;DR

TAAM tackles Graph Class-Incremental Learning without data replay by freezing a GNN backbone and injecting per-task Neural Synapse Modulators to provide task-specific, node-attentive modulation. It introduces Anchored Multi-hop Propagation to generate robust, separable task prototypes for ID inference, ensuring correct NSM selection and stability across unknown tasks. The framework expands a unified classifier only for new classes while training minimal NSM parameters, achieving state-of-the-art AA with zero forgetting on eight datasets and demonstrating strong efficiency. The combination of theoretical grounding and rigorous inductive evaluation positions TAAM as a practical, replay-free solution for continual graph learning with strong stability-plasticity balance.

Abstract

Graph Continual Learning (GCL) aims to solve the challenges of streaming graph data. However, current methods often depend on replay-based strategies, which raise concerns like memory limits and privacy issues, while also struggling to resolve the stability-plasticity dilemma. In this paper, we suggest that lightweight, task-specific modules can effectively guide the reasoning process of a fixed GNN backbone. Based on this idea, we propose Task-Aware Adaptive Modulation (TAAM). The key component of TAAM is its lightweight Neural Synapse Modulators (NSMs). For each new task, a dedicated NSM is trained and then frozen, acting as an "expert module." These modules perform detailed, node-attentive adaptive modulation on the computational flow of a shared GNN backbone. This setup ensures that new knowledge is kept within compact, task-specific modules, naturally preventing catastrophic forgetting without using any data replay. Additionally, to address the important challenge of unknown task IDs in real-world scenarios, we propose and theoretically prove a novel method named Anchored Multi-hop Propagation (AMP). Notably, we find that existing GCL benchmarks have flaws that can cause data leakage and biased evaluations. Therefore, we conduct all experiments in a more rigorous inductive learning scenario. Extensive experiments show that TAAM comprehensively outperforms state-of-the-art methods across eight datasets. Code and Datasets are available at: https://github.com/1iuJT/TAAM_AAMAS2026.

Figures (7)

• Figure 1: Top: A typical CGL scenario where a model must adapt to a stream of new tasks (indicated by new node colors) in an evolving graph. Bottom: An illustration of the Stability-Plasticity Dilemma on the Citeseer benchmark. While strong replay-based methods effectively mitigate the catastrophic forgetting of early tasks, they exhibit diminished plasticity.
• Figure 2: Overview of TAAM framework.TAAM is inserted the frozen GNN backbone,to steer a GNN's reasoning process in stream of graph data.The diagram illustrates the strategy of TAAM.Training Phase: For a new task $\mathcal{G}_k$, a new Neural Synapse Modulator ($\mathrm{NSM}_k$) is created from scratch and trained. Concurrently, its corresponding task prototype $P_k$ is generated and stored.After training is complete, $\mathrm{NSM}_k$ frozen and added to the TAAM module bank.Inference Phase:For incoming test graph $\mathcal{G}^k_t$ with an unknown task ID , a prototype $P^k_t$ generated using AMP,for selects the most relevant "expert" NSM.
• Figure 3: Performance matrices of differernt method on Products dataset.Darker colors indicate higher accuracy. TAAM (Ours) demonstrates consistently high accuracy across all tasks, indicating zero catastrophic forgetting.
• Figure 4: Task ID prediction accuracy of AMP and Laplacian smoothing (LS) of TPP
• Figure 5: Comparison of running time and AA