Computational Performance Aware Benchmarking of Unsupervised Concept Drift Detection

For many AI systems, concept drift detection is crucial to ensure the systems reliability. These systems often have to deal with large amounts of data or react in real time. Thus, drift detectors must meet computational requirements or constraints with a comprehensive performance evaluation. However, so far, the focus of developing drift detectors is on detection quality, e.g.~accuracy, but not on computational performance, such as running time. We show that the previous works consider computational performance only as a secondary objective and do not have a benchmark for such evaluation. Hence, we propose a novel benchmark suite for drift detectors that accounts both detection quality and computational performance to ensure a detector's applicability in various AI systems. In this work, we focus on unsupervised drift detectors that are not restricted to the availability of labeled data and thus being widely applicable. Our benchmark suite supports configurable synthetic and real world data streams. Moreover, it provides means for simulating a machine learning model's output to unify the performance evaluation across different drift detectors. This allows a fair and comprehensive comparison of drift detectors proposed in related work. Our benchmark suite is integrated in the existing framework, Massive Online Analysis (MOA). To evaluate our benchmark suite's capability, we integrate two representative unsupervised drift detectors. Our work enables the scientific community to achieve a baseline for unsupervised drift detectors with respect to both detection quality and computational performance.