Robust Learning Under Distribution Shifts for Non-Stationary Data Environments

Abstract

Machine learning (ML) systems in non-stationary environments, where data distributions vary with time, require robust learning. Conventional deep learning models make the assumption of stationary data, leading to poor performance in the face of concept drift and domain variability. This research suggests a coherent powerful learning model to overcome the change in the distribution of financial transactions. Financial fraud detection dataset, then scaling input data into a uniform range using min-max normalization. The architecture combines adaptive deep learning and distributed learning through an Intelligent Deep Neural Network (AHO-InDNN) that is an Archerfish Hunting Optimizer. It dynamically balances exploration and exploitation while adjusting to evolving fraud patterns. Various types of financial fraud drift, such as abrupt, gradual, and repeat changes, are modeled and identified with a lightweight drift detection module. Incremental learning strategies and online strategies allow real-time adaptation with resource constraints. Moreover, a mechanism of parameter evaluation based on Large Deviation Principle (LDP) is presented to minimize uncertainty and enhance robustness. Experimental results show that the proposed model achieves 98.74% accuracy, 98.42% precision, 98.52% recall, and 98.37% F1-score, outperforming conventional methods. The suggested framework is more stable, generalized, and resilient, offering a practical solution to fraud detection in the dynamic financial non-stationary setting.