Structure-Aware Prototypical Neural Process for Few-Shot Graph Classification.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Science

Background:

• Graph classification is crucial for applications like biological prediction and social analysis.
• Traditional methods struggle with generalization due to handcrafted features.
• Graph Neural Networks (GNNs) show promise but require extensive labeled data, posing challenges for few-shot learning scenarios.

Purpose of the Study:

• To develop a novel approach for few-shot graph classification.
• To address the limitations of existing GNNs in scenarios with limited labeled data.
• To introduce a Structure-aware Prototypical Neural Process (SPNP) model.

Main Methods:

• SPNP utilizes GNNs to capture graph structure information during the encoding stage.
• Structural priors are integrated into latent and deterministic paths for stochastic process representation.
• A novel prototypical decoder with a self-attention mechanism is employed for effective prediction in a defined metric space.

Main Results:

• SPNP demonstrates consistent and significant performance improvements over state-of-the-art methods.
• The model effectively enhances class-level representations, particularly for new classes, via its self-attention mechanism.
• The flexible encoding-decoding architecture allows direct mapping of context samples to predictive distributions.

Conclusions:

• SPNP offers a robust solution for few-shot graph classification.
• The proposed model overcomes the data scarcity limitations of traditional and current GNN approaches.
• SPNP provides a promising direction for future research in efficient and effective graph analysis.