|Zhou Yu, David Gerritsen, Amy Ogan, Alan Black, Justine Cassell|
In this paper we focus on modeling friendships between humans as a way of working towards technology that can initiate and sustain a lifelong relationship with users. We do this by predicting friendship status in a dyad using a set of automatically harvested verbal and nonverbal features from videos of the interaction of students in a peer tutoring study. We propose a new computational model used to model friendship status in our data, based on a group sparse model (GSM) with L2,1 norm which is designed to accommodate the sparse and noisy properties of the multi-channel features. Our GSM model achieved the best overall performance compared to a non-sparse linear model (NLM) and a regular sparse linear model (SLM), as well as outperforming human raters. Dyadic features, such as number and length of conversational turns and mutual gaze, in addition to low level features such as F0 and gaze at task, were found to be good predictors of friendship status.