Sequential non-stationary dynamic classification with sparse feedback

Many data analysis problems require robust tools for discerning between states or classes in the data. In this paper we consider situations in which the decision boundaries between classes are potentially non-linear and subject to "concept drift" and hence static classifiers fail. The applications for which we present results are characterized by the requirement that robust online decisions be made and by the fact that target labels may be missing, so there is very often no feedback regarding the system's performance. The inherent non-stationarity in the data is tracked using a non-linear dynamic classifier, the parameters of which evolve under an extended Kalman filter framework, derived using a sequential Bayesian-learning paradigm. The method is extended to take into account missing and incorrectly labeled targets and to actively request target labels. The method is shown to work well in simulation as well as when applied to sequential decision problems in medical signal analysis.