Heterogeneous ensemble selection for evolving data streams. (April 2021)

• Record Type: Journal Article
• Authors: Luong, Anh Vu; Nguyen, Tien Thanh; Liew, Alan Wee-Chung; Wang, Shilin
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Data streams -- Heterogeneous ensembles -- Ensemble selection; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107743
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: An online ensemble selection method that takes into account each heterogeneous base classifier's confidence during classification and its overall accuracy on the data stream is proposed. Each base classifier's confidence in classification for a test sample is estimated by a threshold computed dynamically using stochastic gradient descent. The overall accuracy of the base classifier is computed using the prequential accuracy that emphasizes more recent instances in the data stream. Extensive comparative experiments with the state-of-the-art algorithms on online ensemble selection validated the superior performance of our algorithm. Abstract: Ensemble learning has been widely applied to both batch data classification and streaming data classification. For the latter setting, most existing ensemble systems are homogenous, which means they are generated from only one type of learning model. In contrast, by combining several types of different learning models, a heterogeneous ensemble system can achieve greater diversity among its members, which helps to improve its performance. Although heterogeneous ensemble systems have achieved many successes in the batch classification setting, it is not trivial to extend them directly to the data stream setting. In this study, we propose a novel HEterogeneous Ensemble Selection (HEES) method, which dynamically selects an appropriate subset of base classifiers to predict data under the stream setting. We are inspired by the observation that a well-chosen subset of good base classifiers may outperform the whole ensemble system. Here, we define a good candidate as one that expresses not only high predictive performance but also high confidence in its prediction. Our selection process is thus divided into two sub-processes: accurate-candidate selection and confident-candidate selection. We define an accurate candidate in the stream context as a base classifier with high accuracy over the current concept, while a confident candidate as one with a confidence score higher than a certain threshold. In the first sub-process, we employ the prequential accuracy to estimate the performance of a base classifier at a specific time, while in the latter sub-process, we propose a new measure to quantify the predictive confidence and provide a method to learn the threshold incrementally. The final ensemble is formed by taking the intersection of the sets of confident classifiers and accurate classifiers. Experiments on a wide range of data streams show that the proposed method achieves competitive performance with lower running time in comparison to the state-of-the-art online ensemble methods.