Physiological Indicators for User Trust in Machine Learning with Influence Enhanced Fact-Checking

Trustworthy Machine Learning (ML) is one of significant challenges of “black-box” ML for its wide impact on practical applications. This paper investigates the effects of presentation of influence of training data points on machine learning predictions to boost user trust. A framework of fact-checking for boosting user trust is proposed in a predictive decision making scenario to allow users to interactively check the training data points with different influences on the prediction by using parallel coordinates based visualization. This work also investigates the feasibility of physiological signals such as Galvanic Skin Response (GSR) and Blood Volume Pulse (BVP) as indicators for user trust in predictive decision making. A user study found that the presentation of influences of training data points significantly increases the user trust in predictions, but only for training data points with higher influence values under the high model performance condition, where users can justify their actions with more similar facts to the testing data point. The physiological signal analysis showed that GSR and BVP features correlate to user trust under different influence and model performance conditions. These findings suggest that physiological indicators can be integrated into the user interface of AI applications to automatically communicate user trust variations in predictive decision making.