Predicting Disengagement in Short-Term Human-Robot Interactions : A Study in a University Library Setting

Sammanfattning: This report presents a study that investigated the effectiveness of Support Vector Machine (SVM) as a machine learning learning technique for detecting human engagement during short-term interactions with social robots in a library setting. The goal was to build an SVM-based model trained on a dataset of recorded interactions and evaluate its performance in predicting engagement levels. The dataset consisted of 96 recordings of short-term interactions between participants from the KTH library and a robot called Furhat. The dataset was manually labeled to classify each second of the recordings as engaged, disengaged, or null (out of frame). The study employed a portion of the dataset for training the SVM model and tested its performance on the remaining recordings. Different data ratios were used for the training and testing sets (75%-25%, 70%-30%, 60%-40%, and 50%-50%). Feature selection was performed using Sequential Feature Selector (SFS) to identify the most informative features from the dataset. The results showed that the number of features did not significantly impact the SVM classifier’s accuracy. Even a subset of 10 features yielded comparable results to using all 160 features, suggesting that including additional features beyond a certain threshold did not enhance the model’s performance. Both radial basis and polynomial kernels showed similar accuracies across different data rations and feature sets. The study highlights the importance of data quality, the influence of actions units in engagement detection, and the limitations of the dataset’s original purpose that focused on frustration rather than engagement. In conclusion, SVM demonstrated some promise in detecting human disengagement in short-term interactions. However, further research is required to compare SVM with other classifiers and investigate engagement-specific datasets for improved accuracy and generalization.