FAT-TAG: Mitigating Forgetting in Task-Free Temporal Graph Class Incremental Learning

The class Incremental Learning aims to train on task sequences continuously, each task introducing a distinct set of classes. Class Incremental Learning on graphs follows the same paradigm but introduces two additional challenges. First, most existing studies overlook the temporal dimension of real-world graphs. These studies approximate the evolving network by slicing a static graph into ordered tasks and assigning newly added nodes to discrete tasks, yet they fail to model explicit temporal dependencies between nodes. Second, many methods still depend on predefined task boundaries and task identifiers when performing inference. This situation severely restricts their use in genuinely task-free scenarios. In this paper, we introduce FAT-TAG, a task-free Temporal Graph Class Incremental Learning framework. (1) FAT-TAG performs temporal graph tokenization to retain structural dependencies while encoding temporal precedence for Graph Transformer. (2) FAT-TAG adopts a Freeze-and-Adapt Training paradigm that trains the Graph Transformer backbone model only on the initial task and equips every subsequent task with a lightweight Low-Rank adapter, effectively preventing catastrophic forgetting. (3) FAT-TAG incorporates a Time-Aware Gate that routes instance-level samples to the most appropriate LoRA adapter without relying on task identifiers, thus achieving task-free inference. Extensive experiments across four real-world temporal graph benchmarks show that FAT-TAG consistently surpasses state-of-the-art baselines in mitigating forgetting. The source code is available at https://github.com/fengjiyuan/TAG.