Zilean: A modularized framework for large-scale temporal concept drift type classification

Abstract

In the analysis of time series data, particularly in real-world applications, concept drift classification is crucial for enabling models to adapt in a differentiated manner to future data. To address the challenge of identifying diverse types of drift, we propose Zilean, a novel framework that integrates feature-based and predictor-based techniques while accounting for drift residues and fragmentation during repeated drift detection. The framework incorporates the pre-trained BERT-Base language model into its classifier design, leveraging deep learning for automatic drift classification and eliminating the need for judgment curve analysis. To evaluate its performance, experiments were conducted on a variety of real-world and synthetic datasets, each exhibiting different types of concept drift. The results show that on real-world datasets, our framework achieves a classification accuracy of 91.03%, outperforming XGBoost by 7.94% and surpassing TCN-CNN by 4.28%. Additionally, experiments exploring a frozen parameter strategy and the use of a more lightweight language model, DistilBERT, further enhance accuracy to 96.93% and 97.17%, respectively. These findings underscore the framework's effectiveness in large-scale temporal concept drift classification.