Rel-MOSS: Towards Imbalanced Relational Deep Learning on Relational Databases

Researchers introduce Rel-MOSS, a novel graph neural network approach designed to address class imbalance problems in relational database entity classification. The method uses relation-centric gating and minority oversampling techniques to prevent underrepresentation of minority classes, achieving 2-4% performance improvements over existing relational deep learning methods.

The research addresses a critical gap in relational deep learning by tackling the class imbalance problem that has been largely overlooked in prior work. Relational databases inherently contain imbalanced entity distributions—some entity types or categories appear far more frequently than others. Traditional graph neural networks applied to these databases amplify this imbalance, causing models to perform poorly on minority classes while optimizing for majority class accuracy.

Rel-MOSS introduces two key innovations to mitigate this issue. The relation-wise gating controller selectively weights messages from different relationship types, preventing majority-class information from drowning out minority signals during neural message passing. The relation-guided minority synthesizer then generates synthetic minority entities while preserving their relational signatures—ensuring synthetically created examples maintain semantic consistency within the database structure.

This work reflects a maturing discipline in AI infrastructure. As graph neural networks become production tools for enterprise database analytics, practitioners encounter real-world data distributions that rarely achieve class balance. The 2-4% improvement in balanced accuracy and G-Mean metrics may seem modest numerically, but represents the difference between usable and unusable models in domains like fraud detection, rare disease diagnosis, or uncommon transaction classification where minority classes carry disproportionate business value.

The research validates improvements across 12 datasets, suggesting broad applicability rather than optimization for specific cases. For organizations deploying GNN-based systems on relational databases, these findings validate the importance of addressing imbalance at the model architecture level rather than through post-hoc balancing techniques. Future work likely extends these concepts to other graph learning challenges in structured data.

• →Rel-MOSS introduces relation-wise gating to prevent majority classes from overwhelming minority signals in graph neural networks
• →Minority oversampling with relational consistency constraints maintains database semantic integrity during synthetic data generation
• →Testing across 12 datasets demonstrates 2-4% improvements in balanced accuracy metrics for entity classification tasks
• →The work addresses a previously unexplored gap in relational deep learning literature regarding class imbalance handling
• →Production deployment of GNN models on enterprise databases benefits significantly from architecture-level imbalance solutions