Presentation
Machine Learning Considerations for Trust Predictions in HRI: Examination across Labels, Sex Differences, and Sensor Modalities
DescriptionAdvances in AI and robotics increase human-robot interactions, but trust classification via neurophysiological data remains uncertain. Additionally, there is a limited understanding of how sex differences, model choice, data modality, and trust binarization influence trust prediction outcomes. We collected data from thirty-eight participants who interacted with a UR10 robot to perform a planetary gear assembly task under reliable and unreliable robot conditions. The dataset comprises neural activation, functional connectivity, gaze entropy features, and behavioral measures such as trust ratings. Classical ML methods (AdaBoost, XGBoost, Naive Bayes Classifier, Logistic Regression, Random Forest, K-Nearest Neighbors, Multi-Layer Perceptron, Support Vector Machine, and Decision Tree) were evaluated using 5-fold cross-validation with accuracy and F1 score as metrics. Our results show 1) the prediction was better for male (acc = 0.935, mae = 0.065) participants than female participants, 2) random forest performed better across all modalities and sexes (acc = 0.802, mae = 0.198), 3) midpoint-point binarization performed the best (acc = 0.898, mae = 0.102), and 4) fNIRS-HbR data modality led to better prediction (acc = .898, mae = 102). These insights underscore the need to explore trust development across diverse demographics and behaviors, informing effective HRI system design.
Event Type
Lecture
TimeThursday, October 16th5:10pm - 5:30pm CDT
LocationGrand Hall M/N
Augmented Cognition