Mutual Information in the Analysis of Trust Gains from Subsets of Information

Information can increase trust of humans in automated machines. However, assessing the impact of all combinations of information pieces on the trust level of humans might not be practical. This paper assumes that data can be collected from human participants having interacted with some automated machine. We assume a two-stage study in which the participants initially submit their ranking (trust level) when no information is provided, and then provide additional independent rankings for each piece of additional information. The goal is to determine the best combination of information pieces over all combinations without directly asking the participants to rank the possible combinations. The impact of the combinations on the trust ranking is evaluated using the mutual information quantity. We further consider the question of statistical significance in this unique setting, and suggest an optimization objective that examines the trade-off between the impact of the subset on the trust measure, on the one hand, while considering the complexity of the subset, measured by the size of the subset (number of additional pieces of information), on the other hand. We provide a numerical example that shows all aspects discussed in this work.