Analysts aren´t machines: Inferring frustration through visualization interaction

Recent work in visual analytics has explored the extent to which information regarding analyst action and reasoning can be inferred from interaction. However, these methods typically rely on humans instead of automatic extraction techniques. Furthermore, there is little discussion regarding the role of user frustration when interacting with a visual interface. We demonstrate that automatic extraction of user frustration is possible given action-level visualization interaction logs. An experiment is described which collects data that accurately reflects user emotion transitions and corresponding interaction sequences. This data is then used in building HiddenMarkov Models (HMMs) which statistically connect interaction events with frustration. The capabilities of HMMs in predicting user frustration are tested using standard machine learning evaluation methods. The resulting classifier serves as a suitable predictor of user frustration that performs similarly across different users and datasets.