Stream change detection via passive-aggressive classification and Bernoulli CUSUM

Stream change detection (SCD) can be defined as the detection of significant deviations in a continuous stream of data. The need of treating data in an online manner, maintaining the balance between detection and false alarm rate and dealing with data streams of different nature are challenges for a general purpose SCD method. One possible solution to cope with these difficulties is to combine an online one-class classifier with a formal policy to decide whether a significant anomaly has occurred. In this work, a modification of passive-aggressive kernel one-class classification algorithm is proposed and combined with a Bernoulli CUSUM chart to deal with SCD problems. The one-class classifier allows for the characterization of data beyond Euclidean space and the CUSUM chart formally deals with the detection of change points. Experimental results show that this method opens the possibility of tackling practical SCD problems in an effective and accurate way.