Build a Lightweight Dataset and Concept Drift Detection Method for Evolving Time Series Data Streams
Nitin B. Ghatage, Pramod D. Patil · Revue d intelligence artificielle · 2023
Time series forecasting, a potent tool for predicting real-world entities such as financial markets and weather patterns, often grapples with the issue of concept drift, characterized by changes in the behaviour of the time series over time.This study aims to develop a lightweight time series model, efficient in training time, to match the data stream's arrival rate.Furthermore, a method to detect the presence of concept drift in the data stream, regardless of the time point, is discussed.Presented herein is a benchmark dataset, publicly accessible and specifically designed to simulate changing time series scenarios across diverse industries including Energy, Air Quality, and Pollution.This dataset amalgamates synthetic and actual time series along with ground truth concept drift locations, facilitating a comprehensive evaluation of concept drift detection techniques.A novel, lightweight concept drift detection method, which integrates supervised methodologies with statistical metrics to surmount the resource constraints often encountered in streaming data scenarios, is proposed.This method minimizes computational overhead while ensuring reliable drift detection in response to shifting data distributions.Experimental results indicate that the proposed approach surpasses prior methods in computational performance whilst accurately identifying idea drifts in evolving time series data streams.The study contributes a valuable dataset and a lightweight feature selection method, advancing the knowledge in the field of concept drift detection within the context of time series data streams.These advancements provide an efficient technique for tracking changing data patterns across various application domains, thus offering significant implications for future research.