Source-linked AI summary
Graph Prototypical Networks for Few-shot Learning on Attributed Networks
Kaize Ding, Jianling Wang, Jundong Li, Kai Shu, Chenghao Liu, Huan Liu
TL;DR
Few-shot node classification on attributed networks is difficult because many classes have limited labels and graph structure violates assumptions used by standard few-shot learning. GPN combines graph neural network representations, node-informativeness valuation, and episodic meta-learning to build transferable prototypes; experiments on four real-world datasets demonstrate effectiveness against baseline methods.
Problem
Existing node-classification methods rely on sufficient labels across classes, while attributed networks contain long-tailed classes with limited labeled instances.
Method
GPN uses graph neural networks to encode network information and estimate labeled-node informativeness, learning prototypes through episodic semi-supervised meta-training.
Results
Empirical results on four real-world datasets demonstrate GPN’s effectiveness versus baseline methods in few-shot node classification.
Takeaways & Limitations
GPN provides a graph meta-learning framework for generalizing to target few-shot classification tasks on attributed networks.
Abstract
from arXiv · showhide
Attributed networks nowadays are ubiquitous in a myriad of high-impact applications, such as social network analysis, financial fraud detection, and drug discovery. As a central analytical task on attributed networks, node classification has received much attention in the research community. In real-world attributed networks, a large portion of node classes only contain limited labeled instances, rendering a long-tail node class distribution. Existing node classification algorithms are unequipped to handle the \textit{few-shot} node classes. As a remedy, few-shot learning has attracted a surge of attention in the research community. Yet, few-shot node classification remains a challenging problem as we need to address the following questions: (i) How to extract meta-knowledge from an attributed network for few-shot node classification? (ii) How to identify the informativeness of each labeled instance for building a robust and effective model? To answer these questions, in this paper, we propose a graph meta-learning framework -- Graph Prototypical Networks (GPN). By constructing a pool of semi-supervised node classification tasks to mimic the real test environment, GPN is able to perform \textit{meta-learning} on an attributed network and derive a highly generalizable model for handling the target classification task. Extensive experiments demonstrate the superior capability of GPN in few-shot node classification.
1 INTRODUCTION
Attributed networks support important node-classification applications, but long-tailed labels leave many classes with too few labeled nodes for existing methods. GPN addresses this setting through graph-aware meta-learning and instance informativeness estimation.
- Node classification infers missing node labels from a partially labeled attributed network and supports applications including social circle learning, document categorization, and protein classification.
- Few-shot learning can generalize from diverse meta-training tasks to new tasks with only a handful of labeled examples.
- Few-shot learning on attributed networks must address non-i.i.d. node dependencies and unequal informativeness among scarce labeled instances.Ignoring node informativeness increases vulnerability to noise and outliers when labeled data is limited.
- GPN learns a transferable metric space with graph-neural-network encoders and valuators, then predicts labels by comparing nodes with informativeness-weighted class prototypes.Meta-learning across semi-supervised node-classification tasks helps extract network meta-knowledge and improve generalization.
2 RELATED WORK
Related work covers graph neural networks and few-shot learning, including optimization-based and metric-based approaches. Prototypical Networks represent classes through support-set means and classify queries by distance.
- Graph neural networks: Graph neural networks learn node representations by propagating information from neighboring nodes, motivating graph-specific methods for graph-structured data.
- Few-shot learning: Few-shot learning solves new tasks with limited examples by transferring knowledge obtained from previous experiences.
- Few-shot learning: Optimization-based approaches learn parameter updates or initializations for few-shot tasks but typically incur fine-tuning costs.
- Few-shot learning: Metric-based approaches learn matching metrics between query and support sets across tasks.
- Few-shot learning: Prototypical Networks average support examples into class prototypes and classify query instances using Euclidean distances.
3 PROBLEM STATEMENT
Few-shot node classification on attributed networks predicts labels for query nodes from unseen classes using only a few labeled support nodes per class. The central goal is transferable meta-knowledge from training classes.
- Notation: The paper denotes sets with calligraphic fonts, vectors with bold lowercase letters, and matrices with bold uppercase letters.
- Attributed networks: An attributed network is represented by nodes, edges, and node-feature vectors collected in a feature matrix.
- Attributed networks: The network can also be represented as an adjacency matrix and feature matrix, with adjacency entries indicating whether node pairs are connected.
- Few-shot node classification: The task provides substantial labels for training classes and asks the model to predict query labels for disjoint test classes with a few labeled support nodes.
- Few-shot node classification: An N-way K-shot task contains N test classes and K labeled support nodes per class, requiring adaptation to new classes from few examples.
4 GRAPH PROTOTYPICAL NETWORKS
GPN addresses few-shot node classification on attributed networks through episodic semi-supervised meta-learning, graph-based representation learning, and support-node valuation. It forms class prototypes from learned node representations, refines them using node importance, and has edge-linear complexity in practice.
- GPN targets three challenges: meta-learning on non-i.i.d. attributed networks, expressive node representations, and identifying informative labeled nodes.
- 4.1 Episodic Training on Attributed Networks: GPN uses episodic training to mimic testing by sampling diverse tasks from training classes, reducing distribution gaps and supporting generalization to unseen test classes.Each task contains an N-way K-shot support set and a query set sampled from the remaining nodes.
- 4.1 Episodic Training on Attributed Networks: Semi-supervised episodes mask most nodes as unlabeled, allowing GPN to use labeled and unlabeled nodes together with their dependencies during meta-learning.
- 4.2 Network Representation Learning: The network encoder uses stacked GNN layers to aggregate neighborhood information and capture long-range dependencies in low-dimensional node representations.The encoder is denoted f_θ(·) and may use an L-layer GNN architecture.
- 4.3 Node Importance Valuation: Class prototypes are computed from embedded support nodes, while GPN refines them because unweighted means neglect node significance and are sensitive to scarce labeled-data noise.A node valuator estimates importance through graph-based score aggregation, and centrality adjustment contributes additional network information.
- 4.5 Complexity Analysis: GPN’s scoring layer has complexity O(|V|dd′), score aggregation has O(|V| + |E|), and overall complexity is considered linear in edges when |E| ≫ |V|.A GCN layer contributes O(|E|dd′), where d and d′ are input and output feature sizes.
5 EXPERIMENTS
Experiments evaluate GPN on few-shot node classification using multiple attributed-network datasets, baselines, task configurations, and ablation analyses. GPN achieves the best performance across few-shot tasks and remains robust as class size, support size, noise, and query size vary.
- 5.2 General Comparisons: Random-walk and conventional GNN baselines lag on few-shot tasks, while MAML and PN perform poorly because they do not capture node dependencies.GNN models can also overfit when only a small number of labeled instances are available; Meta-GNN improves over other baselines in most cases but declines sharply on Reddit.
- 5.3 Parameter Analysis & Ablation Study: As the test class size increases, performance decreases for all models, but GPN increasingly outperforms GPN-naive and remains more robust.The results attribute GPN’s stronger robustness to its node valuator, which learns more representative class prototypes.
- 5.3 Parameter Analysis & Ablation Study: Increasing support size improves all models, while GPN gains more over GPN-naive with small support sets by estimating each labeled sample’s informativeness.Equal-weight prototype averaging makes GPN-naive more sensitive to noisy data, whereas GPN is reported to be more robust.
6 CONCLUSION
GPN combines graph-based node representations with learned labeled-node informativeness to form transferable class prototypes for few-shot node classification. Meta-learning across semi-supervised tasks that mimic testing supports generalization, and experiments on four real-world datasets demonstrate effectiveness against baselines.
- GPN extracts node representations with multilayer graph neural networks using both node attributes and topological structure.
- A second GNN-based component estimates each labeled node's informativeness for constructing representative class prototypes.
- Query-node labels are computed by measuring similarity with the learned class prototypes in a transferable metric space.
- GPN learns over diverse semi-supervised node-classification tasks designed to mimic the real test environment across many episodes.
- Four real-world datasets demonstrate GPN's effectiveness compared with baseline methods in few-shot node classification.
A.1 Data Accessing and Preprocessing
The experiments use graphs constructed from public data sources, with original-dataset links summarized in Table 4. Amazon-Clothing is one documented graph source and preprocessing example.
- All experimental graphs are constructed from public data sources.
- Table 4 summarizes links for accessing the original datasets used to construct the graphs.
- Amazon-Clothing: The Amazon-Clothing data are truncated to the top-level category “Clothing, Shoes and Jewelry”.