A Pdf-free change detection test for data streams monitoring

IEEE Symposium Series on Computational Intelligence, SSCI 2015

Cape Town

2016.01.07

We experience changes in stationarity/time variance in many practical applications. Since changes modify the operational framework the application is working with, its accuracy performance is in turn affected. When changes can occur, we need to detect them as soon as possible, in general by inspecting features extracted from data, and afterwards intervene to mitigate their effects. In this paper, we propose a novel change detection test based on the least squares density difference estimation. Neither assumptions about the distribution of features are needed, nor the change types are made (the method is pdf-free and can handle arbitrary changes.). What here proposed requires limited data to become operational and thresholds needed to assess the change can be set met to predefined false positive rates. We show through comprehensive experiments the effectiveness of the detection method and point out how it outperforms other related methods.