Concept Drift Detection via Boundary Shrinking

Abstract

A change in data distribution, called “concept drift,” often degrades performance of machine-learning models trained beforehand. However, detecting concept drift and identifying where it occurs by hand is a quite costly task and it sometimes overlooks drift until the performance is significantly degraded during operation. In this study, we present an unsupervised concept-drift-detection method with an ensemble of inspector models called “Concept-Drift Detection via Boundary Shrinking (CDDBS)”. To reduce delays in drift detection, we train the inspector models used in CDDBS so that their decision boundaries would be intentionally shrunk in a classification region of a certain class. These shrunk decision boundaries also make it possible to identify where the drift occurs without using true labels because they react individually and sensitively to drift occurring in their corresponding classes. We evaluated the effectiveness of CDDBS by using a simple numerical dataset, several public synthetic benchmark datasets and high-dimensional real images in the CIFAR-10 dataset with synthetic drift. CDDBS outperformed other drift-detection methods in sensitivity of drift detection without increasing the number of false alarms and successfully identified drift-occurring classes without using true labels.