Adaptive Prototype-Guided Personalized Propagation for Heterophilic Graphs With Missing Data.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Science

Background:

• Graph neural networks (GNNs) excel on homophilic graphs but falter with heterophily and missing node attributes.
• Heterophily causes neighborhood semantic inconsistency, while missing features obscure node identity, creating the heterophily-missing coupling (HMC) problem.
• HMC reduces information reliability and breaks standard message propagation assumptions in graphs.

Purpose of the Study:

• To propose a novel framework, Adaptive Prototype-guided Personalized Propagation (APP), to address the challenges of heterophily and feature missingness in graph data.
• To enhance node classification performance on graphs exhibiting the heterophily-missing coupling (HMC).

Main Methods:

• Semantic Rectification via Prototypes (SRPs): Aligns neighborhood information with prototype semantics to reduce noise.
• Personalized Virtual Propagation (PVP): Uses clustering to create virtual edges for effective feature imputation by minimizing Dirichlet energy.
• Adaptive Representation Synergy (ARS): Consolidates features using prototype-guided weighting and contrastive learning for improved representation quality.

Main Results:

• The APP framework consistently improves node classification performance on heterophilic graphs with missing features.
• Achieved up to 11.22% performance improvement over state-of-the-art baselines.
• Significantly reduced feature imputation error.

Conclusions:

• The proposed APP framework effectively tackles the heterophily-missing coupling (HMC) problem in graph neural networks.
• APP offers a robust solution for node classification in challenging graph environments with both heterophily and missing attributes.
• The framework demonstrates superior performance and imputation accuracy compared to existing methods.