SELF-ADAPTIVE MACHINE LEARNING MODELS IN INTELLIGENT INFORMATION SYSTEMS

Abstract

This study investigates adaptive machine learning strategies designed to maintain reliable analytics under conditions of concept drift in continuously evolving data streams. In modern intelligent information systems, streaming data environments frequently experience distributional changes that can significantly degrade the predictive performance of traditional static models. To address this challenge, the study systematically compares several adaptive learning paradigms, including static models, online learning approaches, drift-aware algorithms, and deep adaptive architectures. Controlled streaming datasets with explicitly modelled drift events are used to simulate realistic dynamic environments and evaluate the behaviour of different strategies under changing data distributions. A unified experimental framework is developed to assess multiple performance dimensions, including predictive accuracy, robustness to drift, recovery dynamics after drift events, temporal stability of predictions, and computational resource efficiency. Experimental results demonstrate that static models are highly sensitive to distributional shifts, exhibiting rapid performance degradation and limited ability to recover after drift occurs. Online learning methods provide partial adaptation to changing patterns, while drift-aware algorithms and deep adaptive models consistently show superior resilience, achieving faster recovery, lower degradation, and improved prediction stability over time. In addition, system-level analysis highlights the importance of balancing predictive performance with latency and computational constraints when designing adaptive analytics solutions. The findings confirm that explicit drift detection, adaptive model updating, and robust learning strategies are essential for sustaining reliable real-time analytics in dynamic streaming environments and intelligent information systems.