Unsupervised Concept Drift Detection Based on Stacked Autoencoder and Page-Hinckley Test

Data streams are often subject to concept drift, which can gradually reduce the reliability of learning models over time in data stream mining. To maintain model accuracy and enhance its robustness, it is crucial to detect concept drift and update the learning model accordingly. The majority of drift detection methods rely on the assumption that true labels are immediately available, which is challenging to implement in real-world scenarios. Therefore, it is more practicable to detect concept drift in an unsupervised manner. This paper proposes an unsupervised Drift Detection method based on Stacked Autoencoder and Page-Hinckley test (DD-SAPH). DD-SAPH employs the stacked autoencoder as a medium to represent the distribution of historical data, which extracts hidden features from the reference window. To measure the difference between distributions of historical data and new data, the reconstruction error of the stacked autoencoder on the current window is employed. The Page-Hinckley test dynamically calculates thresholds to warn and alarm concept drift. Experimental results indicate that DD-SAPH outperforms the compared unsupervised algorithms when addressing concept drift on both synthetic and real datasets.