Feature selection for online streaming high-dimensional data: A state-of-the-art review

Knowledge discovery for data streaming requires online feature selection to reduce the complexity of real-world datasets and significantly improve the learning process. This is achieved by selecting highly relevant subsets and minimising irrelevant and redundant features. However, researchers have difficulties in addressing various forms of data. The goal of this article is to present a state-of-the-art review of feature subset selection based on the data form for the high-dimensional data used in online streaming. Through a systematic literature review assessing journal and conference papers from the past five years, detailed discussions on traditional feature selection and online feature selection were presented. Subsequently, a taxonomy of the challenges related to OFS provides a comprehensive review of state-of-the-art OFS and the benchmark methods. Several data forms were identified based on the extensive review: group stream, multi-label, capricious, imbalance, and feature drift. Using critical analysis, the evaluation metrics of online feature selection methods were compared from the perspectives of threshold initialisation, accuracy, high dimensionality, running time, relevancy, and redundancy for the optimal feature subset. An online feature selection framework was derived to illustrate the relationship between the application area, data form, online feature selection methods, evaluation metrics, and tools. Finally, the findings and potential directions for future research were thoroughly discussed. It is suggested that future researchers explore the derived framework and aim to advance each method. (C) 2022 Elsevier B.V. All rights reserved.