Pattern-Aware GNNs Improve Missing Data Handling.

Missing data is a pervasive issue in real-world datasets. Conventional approaches typically either discard incomplete samples or employ imputation techniques, often assuming that data is missing randomly. These methods frequently overlook the potentially valuable information embedded within the patterns of missingness itself. This paper introduces pattern-aware graph neural networks (GNNs) designed to explicitly encode which features are missing, alongside the observed values. The researchers explored four encoding strategies: learned embeddings, frozen random embeddings, statistical features, and hierarchical representations. Evaluations across seven UCI datasets with natural missingness demonstrated substantial improvements, with an average increase of 17% in balanced accuracy and 22% in F1-macro over baseline methods. The benefits varied by dataset, with some showing dramatic gains. Interestingly, even simple random pattern embeddings performed comparably to learned embeddings, suggesting that merely distinguishing between missingness patterns is often more critical than complex optimization of these patterns.