A context-aware approach for handling concept drift in classification

Abstract

Adapting classification models to changes is one of the main challenges associated with learning from data in dynamic environments. In particular, the description of the target concept is not static and may change over time under the influence of varying environmental conditions (i.e. varying context). Although many adaptive learning approaches have been proposed in the literature to address such changes, these are limited in terms of the extent to which the contextual aspects are explicitly identified and utilised. Instead, existing approaches mostly rely on monitoring the effects of drift (in terms of the degradation of the classifier’s performance). Given this, to achieve more effective concept drift management, we propose incorporating context awareness when adapting the classification model to changes. Explicit identification and monitoring of the contextual aspects enable capturing the causes of drift, and hence facilitating more proactive adaptation. In particular, we propose an information-theoretic-based approach for systematic context identification, aiming to learn from data the contextual characteristics of the domain of interest by identifying the context variables contributing to concept changes. Such characteristics are then utilised as important clues guiding the adaptation process of the classification model. Specifically, knowledge of contextual variables are exploited to select the most relevant data for retraining the model via a data weighting model, and to signal the need for data re-selection via a change detection model. The experimental analyses on simulated, benchmark, and real-world datasets, show that such explicit identification and utilisation of contextual information result in a more effective data selection and drift detection strategies, and enable to produce more accurate predictions.