HIVE: Hierarchical Generation of Integrated and Varied Ensembles for Efficient Out-of-Distribution Generalization

Abstract

Out-of-distribution (OOD) generalization in industrial Model-as-a-Service (MaaS) systems is often hindered by a trilemma of heterogeneous distribution shifts, strict inference latency budgets, and rigorous privacy constraints. Traditional robust learning and emerging foundation models frequently struggle with the entangled nature of tabular data or incur prohibitive deployment costs. We propose HIVE, a model-agnostic framework that reframes OOD generalization as the construction of a functionally diverse basis set. Unlike existing methods that pursue a single invariant predictor, HIVE decouples robustness from the model architecture by generating a compact ensemble of experts that enables lightweight and privacy-preserving client-side adaptation. Our primary contributions include: (1) a mechanism-centric clustering approach that identifies latent environments based on predictive behaviors rather than noisy marginal statistics; (2) a hierarchical generation strategy that balances global stability with local specificity; and (3) a diversity--regularized pruning procedure that extracts an efficient subset of models to satisfy real-time industrial requirements. HIVE is uniquely model-free, requiring no access to model internals or gradients, making it compatible with any learning architecture. Extensive offline benchmarking confirms that HIVE consistently outperforms state-of-the-art tabular OOD methods. Since its full-scale deployment in January 2024, HIVE has powered a production fraud detection system where over 95% of clients have opted for this framework over previous standard solutions. Validated on a rigorously siloed dataset of 8 million OOD samples, HIVE demonstrates significant performance lift while strictly satisfying the efficiency mandates of high-throughput industrial environments. Our source code is currently undergoing internal institutional review and will be made publicly available upon completion of the regulatory process at https://github.com/Kee-Qin/code-for-HIVE-for-Efficient-OOD-Generalization.