E-PCDD: An Event-Driven Adaptive Multi-Metric Fusion Framework for Reliable Unsupervised Concept Drift Detection in Evolving Data Streams

Abstract

In real-time data streaming environments, the dynamic and evolving nature of data leads to continuous changes in the underlying data distributions, commonly called concept drift. Traditional drift detection models often struggle to balance detection susceptibility and false alarming ratios. This paper proposes an efficient and optimized hybrid drift detection framework for robust and scalable real-time environments. The proposed technique integrates Jensen–Shannon divergence and Euclidean distance with a sliding window mechanism to capture both probabilistic and geometric distributional changes in the data. An adaptive thresholding strategy based on rolling of statistics dynamically adjusts to evolving data pattern behaviors, eliminating the need for manual parameter tuning or adjustment. In addition to this, an event-based validation method is introduced to reduce redundant detection which ensure stable drift identification over continuous data streams. The framework is implemented on Poker-Hand dataset with synthetic drift injection, dealing with sudden, gradual, and recurring drift. The Experimental results gives strong performance, achieving high accuracy metric score while maintaining low false alarm rate. In general, the proposed framework gives an efficient and practical solution for real-time drift detection problem in applications such as network monitoring and large-scale streaming analytics systems.