Toward a machine learning approach for classifying user goals from user interactions

1 Introduction In this work, we present an evaluation of the utility of collecting and analyzing user-interaction histories to discern information about user intent and user goals. We introduce and compare two techniques for defining the vast space of user goals within a given application. We then offer a supervised, probabilistic approach for mapping series of user interactions into the space of user goals, using collections of goal-classified user-interaction histories as training data. With awareness of user intent, applications may be able to offer tutorial-like advice dynamically, suggesting useful interactions to complete an intended goal. Specifically, we posit that information gained from previous user interactions within a scientific visualization can be leveraged to enhance usability and facilitate scientific discovery: a notion supported by preliminary user-feedback. While interactive visualiza-tion techniques have been shown effective in generating new insight and validating expected hypotheses, we aim to address the lack of attention commonly given to interface usability; an issue that presents a significant roadblock to the widespread adoption of scientific visualizations and workflow systems by domain scientists.