Source-linked AI summary

Meta-GNN: On Few-shot Node Classification in Graph Meta-learning

Fan Zhou, Chengtai Cao, Kunpeng Zhang, Goce Trajcevski, Ting Zhong, Ji Geng

arXiv:1905.09718v1cs.LGstat.ML

TL;DR

Few-shot node classification on graphs is difficult because GNNs must adapt to new classes from very few labeled samples, while graph meta-learning remains underexplored. Meta-GNN trains across many similar few-shot tasks to learn GNN parameter initializations that adapt to new classes, and experiments on three benchmark datasets report improved performance and more general task adaptation.

  • Problem

    Existing GNNs struggle with few-shot node classification because new classes require parameter relearning from very few labeled samples, while meta-learning has seen little application to irregular graph data.

  • Method

    Meta-GNN uses episodic meta-learning across many sampled few-shot tasks to learn GNN parameter initializations that quickly adapt to new classes with few labeled samples.

  • Results

    Meta-GNN improves few-shot node-classification performance over state-of-the-art GNNs on three benchmark datasets and adapts well to new learning tasks.

  • Takeaways & Limitations

    The framework provides a general graph meta-learning approach that can be combined with existing GNNs for few-shot node classification.

Abstract

from arXiv · show

Meta-learning has received a tremendous recent attention as a possible approach for mimicking human intelligence, i.e., acquiring new knowledge and skills with little or even no demonstration. Most of the existing meta-learning methods are proposed to tackle few-shot learning problems such as image and text, in rather Euclidean domain. However, there are very few works applying meta-learning to non-Euclidean domains, and the recently proposed graph neural networks (GNNs) models do not perform effectively on graph few-shot learning problems. Towards this, we propose a novel graph meta-learning framework -- Meta-GNN -- to tackle the few-shot node classification problem in graph meta-learning settings. It obtains the prior knowledge of classifiers by training on many similar few-shot learning tasks and then classifies the nodes from new classes with only few labeled samples. Additionally, Meta-GNN is a general model that can be straightforwardly incorporated into any existing state-of-the-art GNN. Our experiments conducted on three benchmark datasets demonstrate that our proposed approach not only improves the node classification performance by a large margin on few-shot learning problems in meta-learning paradigm, but also learns a more general and flexible model for task adaption.

1 INTRODUCTION

Few-shot node classification remains difficult for GNNs because new classes require parameter relearning from very few labeled nodes. Meta-GNN addresses this gap by applying meta-learning to graph data and is designed as a general framework for adapting GNNs to unseen classes.

  • Research gap: Few-shot node classification challenges GNNs because incorporating new classes requires relearning parameters from very few labeled samples.When each new class has few nodes, existing GNN performance can decline sharply.
  • Research gap: Meta-learning learns transferable knowledge across many tasks and adapts to new tasks using only a few samples.The paper distinguishes meta-training on many tasks from meta-testing on a new task.
  • Research gap: Meta-learning has progressed on Euclidean data such as images and text, but few studies address irregular, noisy, and relational graph data.These graph properties make direct application of existing meta-learning methods difficult.
  • Proposed approach: Meta-GNN is a general graph meta-learning framework for few-shot node classification on graph data, including classes unseen during training.The framework is intended to adapt or generalize to new classes with very few samples.
  • Contributions: The framework can be combined with popular GNN models and improves over state-of-the-art GNNs on three benchmark datasets.These are stated contributions of the paper's evaluation and framework design.

2 METHODOLOGY

Meta-GNN frames few-shot node classification as episodic meta-learning over graph tasks, learning parameters that adapt to new classes from a small support set. It combines GNN representations with task-level updates and evaluates adapted parameters on query nodes.

  • Graph Neural Networks: The GNN jointly uses graph structure and node features through iterative neighborhood aggregation to produce node representations for downstream classification.After l layers, a node representation captures structural information from its l-hop neighbors; aggregation and combination operations define each layer.
  • Meta-Training: Meta-training updates task-adapted parameters from each support set, then optimizes the shared initialization across many tasks.The framework uses one or several gradient-descent steps for each task and performs meta-optimization over the initial parameters rather than the task-specific updated parameters.
  • Meta-Testing: At meta-testing, Meta-GNN adapts to a new task using its support nodes and predicts labels for the query set with the adapted parameters.The new task contains unseen classes, and performance is measured after fine-tuning on a few support samples and evaluating on the query set.
  • Framework Overview: Figure 1 depicts parameter progression from random initialization through one meta-update to the parameters obtained after all meta-updates across M tasks.The black, red, and blue θ values correspond respectively to randomly initialized, once-updated, and fully meta-updated parameters.
  • Task Sampling: Meta-GNN constructs meta-training tasks by sampling classes, support nodes, and query nodes from the training graph.Each task samples |C2| classes, K nodes per class for support, and P remaining nodes for the query set; this process is repeated M times.

3 EXPERIMENTS

Experiments evaluate Meta-GNN on Cora, Citeseer, and Reddit under few-shot node-classification settings, using held-out classes and very small support sets. Meta-GNN improves more strongly as labeled support decreases and performs particularly well on the challenging Reddit dataset, while Meta-SGC and Meta-GCN show dataset-dependent differences.

  • Datasets and setup: Experiments use Cora, Citeseer, and Reddit with modified partitions for few-shot node classification.The source code and datasets are publicly available for reproducibility.
  • Datasets and setup: Cora and Citeseer hold out two classes for meta-testing, while Reddit holds out five classes because it has more unique labels.
  • Datasets and setup: Each support-set class contains only K = 1 or K = 3 samples, and Cora and Citeseer results average accuracy over 50 random support-node selections.For Reddit, the same support nodes are used for each run.
  • Results: Meta-GNN's improvement over baselines on Cora and Citeseer increases as the support set becomes smaller.The reported comparison identifies fewer labeled samples as the setting with greater improvement.
  • Results: On Reddit, Meta-GNN achieves greater improvement, attributed to learning node representations across more tasks and obtaining a more general task-adaptation model.
  • Module comparison: Meta-SGC performs better on Cora, whereas Meta-GCN achieves slightly higher scores on Citeseer and Reddit; SGC-based modules are faster than GCN-based modules.The comparison reports no significant overall discrepancy between the two GNN modules.

4 CONCLUSIONS

The paper concludes that Meta-GNN is a generic framework for few-shot node classification that learns better GNN parameter initialization. It adapts to new tasks and unseen classes with few labeled samples, with encouraging results on three datasets.

  • Meta-GNN uses meta-learning to learn better parameter initialization for GNNs in few-shot node classification.
  • Meta-GNN adapts to new learning tasks and previously unseen classes with few labeled samples.
  • The framework obtains encouraging results on three widely used datasets.
Loading 1905.09718v1…