GINet-DGC: Structural Inductive Biases and Dynamic Generalization Control for High-Dimensional Small-Sample Tabular Data

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Learning from High-Dimensional Low-Sample Size data has long been a persistent challenge in machine learning, because models must infer reliable patterns from limited observations while handling a large number of variables,this challenge is especially common in biomedical applications. High-dimensional small-sample data make predictive modeling vulnerable to unstable training and overfitting. To address this issue, we propose GINet-DGC, an AI framework that integrates feature-wise structural priors with dynamic generalization monitoring. Instead of directly learning an unconstrained first-layer weight matrix, GINet-DGC generates task-specific weights from multi-view feature descriptors, including latent semantic, global distributional, local topological, and hierarchical representations. This structure-constrained weight generation design narrows the feature-interaction search space and acts as an inductive regularizer against noise and redundant molecular features. Furthermore, we introduce an Overfitting-aware Index to monitor the training trajectory and detect the generalization saturation point for adaptive termination. Experiments on eight publicly accessible real-world biomedical HDLSS gene-expression datasets, evaluated under repeated stratified 5×5 cross-validation, show that GINet-DGC achieves competitive and stable performance against 17 baselines. These results suggest that explicitly encoding structural priors can improve robust AI modeling for data-scarce benchmarks.