Pattern-Aware Graph Neural Networks for Handling Missing Data

2026-07-09

AI Summary

This work addresses the limitations of conventional missing data handling methods, which typically assume missingness at random and overlook the informative structure embedded in missing patterns, thereby constraining model performance. The authors propose a novel graph neural network framework that explicitly incorporates missing patterns as a distinct signal and systematically investigates four encoding strategies: learnable embeddings, frozen random embeddings, statistical features, and hierarchical representations. These are integrated with observed values via attention-based or mean aggregation mechanisms to jointly leverage missing structures and available data. Extensive experiments on seven real-world UCI datasets with genuine missingness demonstrate substantial improvements—averaging 17% gains in balanced accuracy and 22% in F1-macro—with the annealing dataset showing an exceptional 80% increase in balanced accuracy, underscoring the critical value of explicitly modeling missing patterns.

Abstract

Missing data is ubiquitous in real-world datasets. Traditional methods either discard incomplete samples or apply imputation techniques that ignore potentially informative missingness patterns, implicitly assuming that missingness occurs randomly. However, missingness patterns might provide additional information. We propose pattern-aware graph neural networks that explicitly encode which features are missing alongside observed values. We used four encoding strategies -- learned embeddings, frozen random embeddings, statistical features, and hierarchical representations -- across seven UCI datasets with naturally occurring missingness. Our Pattern-aware methods achieve substantial improvements over baselines, with an average improvement of 17\% in balanced accuracy and 22\% in F1-macro across all datasets. The benefits vary significantly by dataset: annealing shows dramatic improvement (+80\% balanced accuracy), while hepatitis and soybean show minimal gains (+4--5\%). Notably, even simple random pattern embeddings perform comparably to learned embeddings (0.650 vs 0.663 balanced accuracy), suggesting that distinguishing between patterns may be more important than task-specific optimization. Our ablation study reveals that attention mechanisms, while helpful, are not critical when pattern information is available -- simple mean aggregation with pattern awareness achieves 0.640 balanced accuracy compared to 0.645 for attention-based variants.

Problem

• missing data
• missingness patterns
• graph neural networks
• pattern-aware learning
• imputation

Innovation

• pattern-aware
• graph neural networks
• missing data
• missingness patterns
• embedding strategies