Fast concept drift detection using singular vector decomposition
Dan Shang, Guangquan Zhang, Jie Lü · 2017
Data stream mining is widely used in online applications such as sensor networks, financial transactions, etc. Such systems generate data at high velocity and their underlying distributions may change over time. This is referred to as concept drift problem and it is considered to be the root cause of performance degradation of online machine learning models. To tackle this problem, a reliable and fast drift detection method is required to achieve real time responsiveness to the drifts. This paper presents a fast and accurate drift detection method, namely KS-SVD test - KSSVD, to monitor the distribution changes of the data stream. Our method employs the SVD technique to first check the direction change of the data, followed by a KS test on each direction to detect the univariate distribution changes. Experiments show that our method is efficient and accurate, especially in high dimension situation.