Drift detection and monitoring in non-stationary environments

Detecting changes in data streams is an important area of research in many applications. The challenging issue is to know how to monitor, update and diagnose these changes so that the accuracy of the learner will be improved whatever the nature of the encountered drifts. In this paper a new error distance based approach for drift detection and monitoring, namely EDIST, is proposed. In EDIST, a difference in error distance distributions of two data-windows is monitored through a statistical hypothesis test. The proposed approach is tested using synthetic data and well known real world data sets. Encouraging results were found comparing to others similar approaches. EDIST has reached the best accuracies in most cases and shown more robustness to noise and false alarms.