A Framework to Detect and Classify Time-based Concept Drift
M. Shanthi Thangam, A Bhuvaneswari, J. Sangeetha · Indian Journal of Science and Technology · 2023
Objectives: To design a framework that performs time series decomposition to detect and classify the types of concept drift in a data stream. The aim of this research is to increase the classification accuracy in the detection and classification of drifts. Methods: The proposed method is validated using the Beijing PM2.5 dataset available in the UCI Machine Learning Repository. This dataset has 13 attributes and experiments were performed with the existing drift detection framework algorithms such as EFCDD, Meta-ADD, CIDD, and comparisons were performed with the proposed TBD framework. The outcome of this research is aggregated with Classification accuracy. An effective algorithm selection framework is presented that detects and classifies time-based concept drift existing in the data. The temporal aspects of the data are decomposed to determine the algorithm to be applied to detect and classify the types of drifts. Depending on the decomposed levels, three varied algorithms have been applied and used for the effective detection and classification of time-based drifts. Findings: The performance of the proposed method is validated using the classification accuracy and compared with the existing drift detection framework algorithms. The proposed framework achieves maximum classification accuracy of 95.24% than all the other existing methods. Novelty: A novel framework has been proposed with better classification accuracy for the detection and classification of time-based concept drift. Keywords: Feature Selection, Concept Drift, Multiple Drift Detection, Timeseries decomposition, Classification Accuracy