CAPL: Graph Few-Shot Class-Incremental Learning Via Class-Adaptive Prototype Learning

ICLR 2026 Conference Submission16101 Authors

Keywords: Few-shot class-incremental learning, Graph few shot learning, Graph class-incremental learning

Abstract:

Few-shot class-incremental learning has always been a challenging problem due
to catastrophic forgetting, insufficient labels and class imbalance. Graph few-shot
class-incremental learning(GFSCIL), with the presence of edges between nodes
and complex relationships between classes, further increases the difficulty of the
learning process. Current researches in this field mainly employ meta-learning and
metric-learning approaches. However, these methods do not consider the relation-
ships between classes and treat all classes equally, which does not conform to the
real-world applications. To address these limitations, we propose a class-adaptive
prototype learning (CAPL) method that adaptively processes each class based on
the relationships between classes, thereby alleviating spatial confusion between
new and old classes as well as the catastrophic forgetting problem. Specifically,
we first adopt a class-adaptive spatial reservation module to allocate larger spaces
for the arrival of new classes, preventing confusion between new and old classes.
We then utilize a class-adaptive prototype alignment module for knowledge distil-
lation. By considering the positional relationship between new and old classes in
the feature space, we provide greater flexibility to classes closely related to new
classes while retaining classification information of old classes, thus adapting to
the arrival of new classes. Experiment results demonstrate the superiority of the
proposed method.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 16101