An Error-Based Measure for Concept Drift Detection and Characterization
Résumé
Continual learning is an increasingly studied field, aiming at regulating catastrophic forgetting for online machine learning tasks. In this article, we propose a prediction error measure for continual learning,
to detect concept drift induced from learned data input before the learning step. In addition, we check this measure’s ability for characterization of the drift. For these purposes, we propose an algorithm to compute the
proposed measure on a data stream while also estimating concept drift. Then, we calculate the correlation coefficients between this estimate and our measurement, using time series analysis. To validate our proposal,
we base our experiments on simulated streams of metadata collected from an industrial dataset corresponding to real conversation data. The results show that the proposed measure constitutes a reliable criterion
for concept drift detection. They also show that a characterization of the drift relative to components of the stream is possible thanks to the proposed measure.