Analysis of Feedback Contents and Estimation of Subjective Scores in Social Skills Training.

This paper introduces our analysis results on the feedback contents of Social Skills Training and the consequences of automated score estimation of users' social skills with computational multimodal features. Although previous work showed the possibility of a computerized SST system as a clinical tool, its feedback strategies have not been well-investigated. We focused on the feedback content given by experienced SST trainers in human-human SST sessions to overcome this limitation. We analyzed the points mentioned by experienced SST trainers to determine where they focused during social skills evaluation. We calculated multimodal computational features from video and audio recordings inspired by the results and trained machine learning models for social skills evaluation using these features as input. We trained social skill score prediction models with the highest scores of 0.53 for correlation coefficient and 0.26 for R. Clinical relevance- We described our automated social skills evaluation method with machine learning models toward a computerized SST system, which can be an additional option to boost the effect of SST by experienced trainers in the future.