An Exploratory Study on Group Potency Classification from Non-verbal Social Behaviours

Nicola Corbellini, Eleonora Ceccaldi, Giovanna Varni, Gualtiero Volpe

Abstract: Technological research is increasingly focusing on intelligent computer systems that can elicit collaboration in groups made of a mix of humans and machines. These systems have to devise appropriate strategies of intervention in the joint action, a capability that requires them to be able of sensing group processes such as emergent states. Among those, group potency – i.e., the confidence a group has that it can be effective – has a particular relevance. An intervention targeted at increasing potency can indeed increment the overall performance of the group. As an initial step in this direction, this work addresses automated classification of potency from multimodal group behaviour. Interactions by 16 different groups displaying low or high potency were extracted from 3 already existing datasets: AMI, TA2, and GAME-ON. Logistic Regression, Support Vector Machines, and Random Forests were used for classification. Results show that all the classifiers can predict potency, and that a classifier trained with samples from 2 of the datasets can predict the label of a sample from the third dataset. ...