TAAM:Inductive Graph-Class Incremental Learning with Task-Aware Adaptive Modulation
TAAM leverages lightweight Neural Synapse Modulators and Anchored Multi-hop Propagation for replay-free, resource-efficient continual graph learning, outperforming SOTA with zero forgetting.
Key Findings
Methodology
TAAM integrates a frozen Simple Graph Convolution (SGC) backbone with task-specific Neural Synapse Modulators (NSMs). Each task trains a dedicated NSM, which modulates the backbone’s output via node-attentive FiLM parameters. AMP generates robust task prototypes through multi-hop diffusion based on Approximate Personalized PageRank, enabling accurate task ID inference without labels. During training, only NSMs and the classifier are updated, ensuring parameter efficiency and preventing catastrophic forgetting. The system dynamically selects the appropriate NSM at inference based on prototype similarity, maintaining knowledge integrity across tasks.
Key Results
- On eight benchmark datasets, TAAM achieves the highest average accuracy and zero catastrophic forgetting (AF=0%). For example, on Citeseer, it reaches 85.2% AA, surpassing all baselines. On Products, it attains 94.5%, outperforming existing methods by over 10%. Ablation studies confirm the importance of NSM and AMP modules. The results demonstrate that TAAM effectively balances stability and plasticity, with significant improvements over replay-based and parameter-isolation approaches, especially in inductive settings where data leakage is prevented.
Significance
This work addresses critical limitations of existing continual graph learning methods, notably the reliance on large pre-trained models and data replay. By proposing a resource-efficient, replay-free framework with robust task inference, it advances the field towards practical, privacy-preserving solutions. Its ability to prevent catastrophic forgetting while maintaining high plasticity opens new avenues for real-world applications such as dynamic social networks, evolving knowledge graphs, and adaptive recommendation systems, where data privacy and model efficiency are paramount.
Technical Contribution
The paper introduces a novel combination of lightweight, node-attentive modulators with a theoretically grounded AMP task ID inference mechanism. This design allows for fine-grained, task-specific adaptation without altering the backbone, and guarantees high task discrimination accuracy. The framework’s theoretical analysis of AMP’s robustness and the integration of parameter-efficient modules distinguish it from prior methods relying on heavy pre-training or replay buffers, providing a new paradigm for scalable continual graph learning.
Novelty
This is the first work to eliminate dependence on pre-trained models in graph continual learning by employing lightweight modulators and multi-hop diffusion-based prototypes. Unlike existing prompt-based or replay methods, TAAM achieves high performance with minimal parameters, offering a flexible, privacy-preserving, and scalable solution. Its innovative task inference mechanism and modular design set a new benchmark in the field.
Limitations
- TAAM’s performance may degrade on highly complex or heterogeneous graphs where neighborhood information is insufficient for accurate task ID inference. The reliance on neighborhood diffusion could be less effective in sparse or noisy graphs. Additionally, the current design assumes a static backbone, limiting adaptability to evolving architectures. Future work should explore dynamic backbone integration and robustness in more diverse graph types.
Future Work
Future directions include extending TAAM to dynamic and heterogeneous graphs, integrating self-supervised signals for better task inference, and scaling to larger, real-world datasets. Developing adaptive backbone architectures and exploring multi-task joint learning are also promising avenues to enhance the framework’s flexibility and applicability in complex environments.
AI Executive Summary
Graph Neural Networks (GNNs) have revolutionized static graph analysis, but real-world data often evolves over time, posing significant challenges for continual learning. Traditional solutions rely heavily on data replay or large pre-trained models, which raise privacy concerns, demand substantial storage, and limit flexibility. These approaches also struggle with the stability-plasticity dilemma—balancing knowledge retention with new learning—especially without explicit task labels during inference.
To address these issues, this paper introduces TAAM, a novel framework that combines a fixed GNN backbone with lightweight, task-specific Neural Synapse Modulators (NSMs). Each task trains a dedicated NSM, which modulates the backbone’s internal representations through node-attentive FiLM parameters. This design ensures that task-specific knowledge is encapsulated within small modules, preventing catastrophic forgetting without replay. A key innovation is the Anchored Multi-hop Propagation (AMP) method, which generates robust task prototypes by diffusing node features across multiple hops, inspired by Personalized PageRank. During inference, AMP accurately identifies the task of incoming data by comparing prototype similarities, enabling the system to select the correct NSM.
Extensive experiments across eight benchmark datasets demonstrate TAAM’s superiority over state-of-the-art methods, achieving the highest accuracy and zero forgetting. For example, on Citeseer, it reaches 85.2% AA, surpassing existing approaches by significant margins. The framework’s efficiency, combined with its theoretical guarantees for AMP’s robustness, makes it highly promising for practical deployment. This work paves the way for privacy-preserving, scalable, and adaptive graph learning in dynamic environments, with broad implications for social networks, knowledge graphs, and beyond. Future work will focus on extending to more complex graph types and larger-scale applications, further enhancing its real-world impact.
Deep Dive
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.