[Paper Review] Structure-Enhanced Meta-Learning For Few-Shot Graph Classification

Shunyu Jiang, Fuli Feng+3 more|arXiv (Cornell University)|Mar 5, 2021

Advanced Graph Neural Networks24 references4 citations

TL;DR

This paper proposes SMF-GIN, a metric-based meta-learning framework for few-shot graph classification that explicitly models both global and local graph structures using adaptive aggregation and substructure attention. Evaluated on Chembl and TRIANGLES, SMF-GIN achieves state-of-the-art performance by enhancing representation learning through structure-aware distance metrics.

ABSTRACT

Graph classification is a highly impactful task that plays a crucial role in a myriad of real-world applications such as molecular property prediction and protein function prediction.Aiming to handle the new classes with limited labeled graphs, few-shot graph classification has become a bridge of existing graph classification solutions and practical usage.This work explores the potential of metric-based meta-learning for solving few-shot graph classification.We highlight the importance of considering structural characteristics in the solution and propose a novel framework which explicitly considers global structure and local structure of the input graph. An implementation upon GIN, named SMF-GIN, is tested on two datasets, Chembl and TRIANGLES, where extensive experiments validate the effectiveness of the proposed method. The Chembl is constructed to fill in the gap of lacking large-scale benchmark for few-shot graph classification evaluation, which is released together with the implementation of SMF-GIN at: https://github.com/jiangshunyu/SMF-GIN.

Motivation & Objective

• Address the challenge of few-shot graph classification where labeled graphs per class are extremely limited, common in molecular and biological applications.
• Overcome the limitation of existing meta-learning methods that fail to explicitly leverage graph structural characteristics such as global topology and local substructures.
• Develop a transparent, structure-enhanced meta-learning framework that improves generalization to novel graph classes with minimal labeled examples.
• Introduce a large-scale benchmark dataset, Chembl, to fill the gap in few-shot graph classification evaluation.
• Demonstrate the superiority of structure-aware representation learning over standard GNNs and existing meta-learning baselines in low-shot settings.

Proposed method

• Employ Graph Isomorphism Network (GIN) as the backbone encoder to learn graph representations at multiple GNN layers.
• Design an adaptive aggregation module that assigns layer-specific weights to GNN representations based on graph-specific structural properties, enhancing global structure modeling.
• Introduce a substructure attention mechanism to extract and fuse representative local substructures (e.g., molecular scaffolds) into the overall graph representation.
• Combine global and local representations through a learnable fusion mechanism to improve the metric space for few-shot classification.
• Train the framework using a metric-based meta-learning objective, where support sets are used to compute distance-based predictions in the learned embedding space.
• Utilize a multi-task learning setup with episode-based training, where each episode simulates a new N-way K-shot classification task.

Experimental results

Research questions

• RQ1 Can explicit modeling of both global and local graph structures improve few-shot graph classification performance?
• RQ2 How does adaptive aggregation of multi-layer GNN representations affect generalization in low-shot settings?
• RQ3 To what extent does incorporating substructure representations enhance the discriminative power of graph embeddings?
• RQ4 Does a structure-aware metric-based meta-learning framework outperform standard gradient-based or non-structure-aware meta-learning methods on graph classification?
• RQ5 How effective is the proposed method on large-scale, real-world few-shot graph benchmarks like Chembl?

Key findings

• SMF-GIN achieves state-of-the-art performance on the TRIANGLES and Chembl datasets for few-shot graph classification, demonstrating superior generalization with limited labeled graphs.
• The ablation study confirms that both global structure modeling via adaptive aggregation and local structure modeling via substructure attention significantly improve classification accuracy.
• The proposed framework outperforms vanilla GIN and existing meta-learning baselines, especially in low-shot regimes (e.g., 1- or 2-shot settings).
• The release of the Chembl benchmark dataset provides a large-scale, realistic evaluation suite for future few-shot graph learning research.
• The attention mechanisms in SMF-GIN are interpretable, revealing that substructures like molecular scaffolds play a decisive role in classification decisions.
• The framework shows robustness across diverse graph types, including molecules with varying diameters and structural complexity, due to its multi-granularity structure modeling.