Social relationships, such as interpersonal closeness or rapport, can lead to improved student learning, but such dynamic, interpersonal phenomena can be difficult for educational support technologies to detect. In this paper, we describe an approach for rapport detection in peer tutoring, using temporal association rules learned from nonverbal, social, and on-task verbal behaviors. From a corpus of 60 hours of annotated multimodal peer tutoring data, we learn the temporal association between behaviors and the rapport score for each 30-second "thin-slice". We then train a stacked ensemble classification model on those association rules and evaluate our ability to reliably predict rapport using multimodal behavioral data. We find that our approach allows us to predict rapport well above chance, and more accurately than two baseline models. We are able to predict high rapport more accurately for strangers and low rapport more accurately for friends, which we believe holds promise for the integration of rapport detection into collaborative learning supports and intelligent tutoring systems. [For the full proceedings, see ED596512.]