Instance-Prototype Affinity Learning for Non-Exemplar Continual Graph Learning

doi:10.1016/j.patcog.2026.114911

Instance-Prototype Affinity Learning for Non-Exemplar Continual Graph Learning

2026-09-17

0

PRE

AI

DOI：10.1016/j.patcog.2026.114911

Original

Original request for help

Save

Abstract

En 中文

• Empirical and theoretical studies indicate Prototype Contrastive Learning induces less drift than Prototype Replay. • We propose a novel PCL-based NECGL paradigm effectively balancing stability and plasticity. • We integrate graph topology into prototype computation to boost learning capacity. • We propose a flexible distillation method that mitigates catastrophic forgetting while preserving model plasticity. • We include hard examples in PCL to improve inter-class separability.

Keywords:

Continual graph learning

Non-exemplar

Prototype contrastive learning

Node classification

Instance-Prototype Affinity Learning for Non-Exemplar Continual Graph Learning

Abstract

Journal

Organization

Cited Papers