Models and Model Biases for Automatically Learning Task Switching Behavior

Machine learning techniques have been applied to several kinds of human data including speech recognition and goal or user identification. When learning on such data, it is important to use models that are not strongly biased against properties of the data, or the variable assignments learned may be largely incorrect. We are working with data sources for user interface event data and examining the applicability of dynamic Bayesian networks (DBNs) to context tracking. Specifically, we identify the value and transition points of a hidden task variable; this problem is known as segmentation. Our data is drawn from command line interaction collected in a real work setting and window event traces taken during a controlled behavioral study. We have applied discrete time hidden Markov models (HMMs) and DBNs to these data sets, but these methods are fundamentally Markovian and, as a result, cannot correctly learn the properties of hidden variables with nongeometrically distributed dwell times. We believe that using semi-Markov models may better capture some underlying structure and allow for better segmentation. In this paper, we describe the experimental protocols performed, examine the bias of typical DBNs and HMMs towards geometric variable dwell times, and assess the validity of this assumption. We discuss the issues of applying semi-Markov DBNs to the available data.