SLED: Semi-supervised Locally-weighted Ensemble Detector

Abstract

Concept drift detection refers to the process of detecting changes in the underlying distribution of data. Interest in the data stream mining community has increased, because of their role in improving the performance of online learning algorithms. Over the years, a myriad of drift detection methods have been proposed. However, most of these methods are single detectors, which usually work well only with a single type of drift. In this research, we propose a semi-supervised locally-weighted ensemble detector (SLED), where the relative performance among its base detectors is characterized by a set of weights learned in a semi-supervised manner. The aim of this technique is to effectively deal with both abrupt and gradual concept drifts. In our experiments, SLED is configured with ten well-known drift detectors. To evaluate the performance of SLED, we compare it with single detectors as well as state-of-the-art ensemble methods on both synthetic and real-world datasets using different performance measures. The experimental results show that SLED has fewer false positives, higher precision, and higher Matthews correlation coefficient while maintaining reasonably good performance for other measures.