Abstract:Node classification with Graph Neural Networks (GNN) under a fixed set of labels is well known in contrast to Graph Few-Shot Class Incremental Learning (GFSCIL), which involves learning a GNN classifier as graph nodes and classes growing over time sporadically. We introduce inductive GFSCIL that continually learns novel classes with newly emerging nodes while maintaining performance on old classes without accessing previous data. This addresses the practical concern of transductive GFSCIL, which requires storing the entire graph with historical data. Compared to the transductive GFSCIL, the inductive setting exacerbates catastrophic forgetting due to inaccessible previous data during incremental training, in addition to overfitting issue caused by label sparsity. Thus, we propose a novel method, called Topology-based class Augmentation and Prototype calibration (TAP). To be specific, it first creates a triple-branch multi-topology class augmentation method to enhance model generalization ability. As each incremental session receives a disjoint subgraph with nodes of novel classes, the multi-topology class augmentation method helps replicate such a setting in the base session to boost backbone versatility. In incremental learning, given the limited number of novel class samples, we propose an iterative prototype calibration to improve the separation of class prototypes. Furthermore, as backbone fine-tuning poses the feature distribution drift, prototypes of old classes start failing over time, we propose the prototype shift method for old classes to compensate for the drift. We showcase the proposed method on four datasets.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the **few - shot classification problem of graph neural networks (GNNs) in the incremental learning environment**, specifically **Inductive Graph Few - shot Class Incremental Learning (Inductive GFSCIL)**. Different from traditional node classification tasks, GFSCIL involves the continuous growth of nodes and classes in the graph over time, and the model needs to keep learning new classes in this dynamic environment while maintaining performance on old classes. #### Specific problem description: 1. **Inductive vs. Transductive settings**: - **Transductive GFSCIL**: It is assumed that the model can access the historical data of the entire graph, which is not always feasible in practical applications. - **Inductive GFSCIL**: It does not store past sub - graphs and is trained only based on the current sub - graph, which better meets the requirements of privacy protection and memory limitations in real - world scenarios. 2. **Catastrophic Forgetting**: - During the incremental learning process, because the model cannot access past data, it may forget the knowledge of old classes, resulting in a performance decline. 3. **Overfitting problem**: - New classes usually have only a small number of labeled nodes, which easily leads to overfitting. 4. **Prototype Drift**: - As the model parameters are updated, the feature distribution of old classes will change, causing the prototypes of old classes to be no longer accurate. ### Solution: To address the above challenges, the authors propose the **Topology - based class Augmentation and Prototype calibration (TAP)** method, which mainly includes the following parts: 1. **Triple - branch Multi - topology Class Augmentation (TMCA)**: - **Full Topology**: Use the complete base graph. - **Topology - Free**: Remove all edges and only keep the node features. - **Topology - Varying**: Divide the base class into multiple subsets and disconnect the connections between different subsets to increase structural diversity. 2. **Iterative Prototype Calibration (IPCN)**: - Use the query set to iteratively calibrate the prototypes of new classes, similar to the k - means algorithm, to improve the representativeness of new class prototypes. 3. **Prototype Shift Update (PSO)**: - Adjust the prototypes of old classes by estimating feature changes to compensate for the feature distribution drift caused by the update of model parameters. ### Summary: By introducing the TAP method, this paper solves the problems of catastrophic forgetting, overfitting, and prototype drift in Inductive GFSCIL, enabling the model to effectively learn new classes and maintain performance on old classes without accessing past data.

Inductive Graph Few-shot Class Incremental Learning

Two-level Graph Network for Few-Shot Class-Incremental Learning

Geometer: Graph Few-Shot Class-Incremental Learning Via Prototype Representation

Constructing Sample-to-Class Graph for Few-Shot Class-Incremental Learning

Incremental Graph Classification by Class Prototype Construction and Augmentation

Few-Shot Incremental Learning with Continually Evolved Classifiers

Hybrid Graph Neural Networks for Few-Shot Learning

A Cognition-Driven Framework for Few-Shot Class-Incremental Learning

Pseudo Initialization Based Few-Shot Class Incremental Learning

A survey on few-shot class-incremental learning

Few-Shot Incremental Learning for Label-to-Image Translation

InfoMax Classification-Enhanced Learnable Network for Few-Shot Node Classification

Dynamic Support Network for Few-Shot Class Incremental Learning

Few Shot Class Incremental Learning via Grassmann Manifold and Information Entropy

Few-Shot Class-Incremental Learning Via Class-Aware Bilateral Distillation

Supplementary Material for Few-Shot Class-Incremental Learning

Analogical Learning-Based Few-Shot Class-Incremental Learning

Coarse-To-Fine Incremental Few-Shot Learning

Knowledge Adaptation Network for Few-Shot Class-Incremental Learning

Learning to Classify With Incremental New Class

Towards Robust Graph Incremental Learning on Evolving Graphs