A just-in-time adaptive classification system based on the intersection of confidence intervals rule

Classification systems meant to operate in nonstationary environments are requested to adapt when the process generating the observed data changes. A straightforward form of adaptation implementing the instance selection approach suggests releasing the obsolete data onto which the classifier is configured by replacing it with novel samples before retraining. In this direction, we propose an adaptive classifier based on the intersection of confidence intervals rule for detecting a possible change in the process generating the data as well as identifying the new data to be used to configure the classifier. A key point of the research is that no assumptions are made about the distribution of the process generating the data. Experimental results show that the proposed adaptive classification system is particularly effective in situations where the process is subject to abrupt changes.

► We propose a novel just-in-time adaptive classifier to operate in nonstationary environments.
► Process stationarity is monitored by a change-detection test based on the intersection of confidence intervals rule.
► An adaptive knowledge-base management of the classifier follows any detection.
► Only training samples coherent with the current state of the process are considered.
► The classifier effectively and promptly adapts to sequences of abrupt changes.