Enhanced self-directed training (ESDT): dynamic data balancing in regression models

Abstract

Enhanced Self-Directed Training (ESDT) is a novel framework designed to enhance regression-based Multilayer Perceptron (MLP) models through strategic synthetic data augmentation. Unlike traditional methods that rely on random data generation, ESDT focuses on high-loss validation cases, applying Synthetic Data Extension (SDE) techniques, such as SMOTE-based interpolation for continuous data and time-based translation for discontinuous events. This strategy significantly improves model generalization, reduces overfitting, and boosts performance in unbalanced datasets. By tailoring data augmentation efforts to specific areas where the model encounters challenges, ESDT effectively increases data diversity and resilience, ensuring robust predictive modeling across various complex domains. This research establishes ESDT as a robust, adaptive training methodology, promising advancements in scalability and efficiency for future machine learning applications.