Adaptive Probabilistic Classification of Dynamic Processes: A Case Study on Human Trust in Automation

Classification algorithms have traditionally been developed based on the assumption of independent data samples characterized by a stationary distribution. However, some data types, including human-subject data, typically do not satisfy the aforementioned assumptions. This is relevant given the growing need for models of human behavior (as they relate to research in human-machine interaction). In this paper, we propose an adaptive probabilistic classification algorithm using a generative model. We model the prior probabilities using a Markov decision process to incorporate temporal dynamics. The conditional probabilities use an adaptive Bayes quadratic discriminant analysis classifier. We implement and validate the proposed algorithm for prediction of human trust in automation using electroencephalography (EEG) and behavioral response data. An improved accuracy is obtained for the proposed classifier as compared to an adaptive classifier that does not consider the temporal dynamics of the process being considered. The proposed algorithm can be used for classification of other human behaviors measured using psychophysiological data and behavioral responses, as well as other dynamic processes characterized by data with non-stationary distributions.