RAFE-Net: A Residual Adaptive Residual Ensemble Feedback Network to Improve Accuracy of Prediction

Abstract

Prediction systems are the most important part of making decisions based on data. They affect areas like finance, healthcare, industrial automation, and climate science. Even with improvements in deep learning and machine learning, current predictive models still have three big problems: they make more mistakes with each prediction, they are sensitive to changes in the data distributions they are based on, and they can't always capture how different features interact with each other. This paper presents RAFE-Net (Residual Adaptive Feedback Ensemble Network), an innovative algorithm specifically formulated to tackle these constraints. RAFE-Net uses the best parts of ensemble learning, a residual feedback module (RFM) that learns from mistakes all the time, and a distribution shift detector (DSD) that changes predictions when data distributions change a lot. The framework not only makes things more accurate, but it also makes them more robust and easier to understand. This makes it good for high-stakes situations like fraud detection and medical diagnosis. Experimental assessments utilising benchmark datasets—such as the M4 time series dataset, the IEEE-CIS fraud detection dataset, and various datasets from the UCI Machine Learning Repository—illustrate that RAFE-Net attains a predictive accuracy improvement of up to 7.2% and a reduction in false positives by 12.5% relative to leading-edge baselines. These findings underscore the promise of feedback-driven ensemble frameworks as the forthcoming generation of predictive modelling systems.

Keywords: Prediction accuracy, residual learning, ensemble models, adaptive feedback, time series analysis, anomaly detection.