Mitigating Structural Overfitting: A Distribution-Aware Rectification Framework for Missing Feature Imputation

A team of researchers has published a paper introducing the DART (Distribution-Aware Rectification) framework, a novel solution to a critical flaw in graph neural networks (GNNs) called 'structural overfitting.' This problem occurs when models trained to fill in missing data (imputation) rely too heavily on the specific connections in their training graph, causing them to fail on disjointed graphs or new, unseen data—a common issue in real-world applications like user profiling and cold-start recommendations.

Current state-of-the-art methods use diffusion-based techniques, like feature propagation, which smooth data across a graph's structure. The DART framework innovates by adding three key components. First, a Global Structural Augmentation (GSA) module establishes correlations to connect disjoint parts of a graph. Second, a semantic rectifier, based on masked autoencoding, learns the underlying data manifold to recover authentic details lost by over-smoothing. Finally, a test-time distribution rectification mechanism projects predictions back onto this learned manifold during inference, closing the 'inductive gap' when applying the model to new structures.

The research is notable for its practical validation. The team created a new benchmark dataset called 'Sailing,' derived from real voyage records with naturally missing attributes, arguing that synthetically masked data doesn't reflect true sparsity patterns. Extensive testing on Sailing and six public datasets showed DART significantly outperforming existing methods in both transductive (same graph) and inductive (new graph) settings. The code and dataset have been made publicly available, and the paper is accepted for SIGIR 2026.