Endowing machines with social intelligence is a fundamental goal of artificial social intelligence. Dealing with human-centered phenomena requires, however, a considerable amount of manually annotated data, making data annotation a costly and challenging task that hinders the training of supervised learning algorithms. In this study, we apply an approach grounded on Graph Convolutional Network (GCN) to alleviate the annotation burden. As a test bed, we select emergent states analysis with specific reference to the team potency. At first, we build the POTENCY dataset by fusing three datasets on social interaction. Next, we compute a set of multimodal features characterizing the social behavior of the team members and the team as one. Finally, we feed the POTENCY dataset to a semi-supervised GCN, trained on a binary node classification task, with variable amounts of labels. We show that GCN can assign team potency labels to an unlabeled team in the dataset by using only a few labeled examples (i.e., 10% of data), with performances comparable to or higher than those of two baseline algorithms carrying out the same task in a fully supervised way.