A Data-Driven Framework for Detecting and Mitigating Concept Drift in Adaptive artificial Neural Networks

Learning in non-stationary environment is a daunting task as the distributions and their characteristics change or evolve over time and often succumb to a problem known as Model decay due to Concept drift. Concept drift hinders performance of the Machine learning models. Traditionally many Machine learning architectures like Random forest, Decision Trees and various ensemble techniques are employed as learners to capture and adapt to Concept drift. The presented work aims at examining the performance of artificial Neural Networks in Non-Stationary and continuously evolving environments. The experimentation uses Kolmogorov-Smirnov test to detect drift points as a drift detector with learner as artificial Neural Networks and the adaptation strategy of retraining if drift occurs due to change in the statistical distributions of data. The experimentation involved Random forest and Decision Trees as a base model and the performance of artificial Neural Networks is compared against the base architectures. The artificial Neural Networks performed better in terms of precision of 0.657 but with a bad recall of 0.398 indicating some misclassification for true positives. Comparatively Random forest yielded good performance in terms of precision and recall of 0.557 and 0.547 respectively, with Decision Tree learner performed poorly on the drifting dataset. The results underscore the need for more robust drift detection and adaptation mechanisms. Future work will focus on integrating advanced techniques such as online learning and ensemble methods to improve the model’s ability to handle concept drift and maintain predictive performance in dynamic environments.