Collaborative problem solving behaviors are difficult to identify and foster due to their amorphous and dynamic nature. In this paper, we investigate the value of considering early class period behaviors, based on small group development theory, for building predictive machine learning models of collaborative behaviors during problem solving. Over 12 weeks, 20 small groups of undergraduate students solved problems facilitated by a digital joint problem space tool on tablet computers, in the 50-minute discussion component of an engineering course. We annotated 16,270 video clips of groups for collaborative behaviors including task relatedness, talk content, peer interaction, teaching assistant interaction, and tablet usage. We engineered two subsets of features from tablet log file data: onset features (early collaborative problem solving behavior characteristics calculated from the first ten minutes of the class) and concurrent features (more general collaborative behaviors from the whole class period). We compared accuracy between the onset, concurrent, and onset + concurrent features in machine learning models. Results exhibited a U-shaped pattern of accuracy over class time, and showed that onset features alone could not be used to effectively model groups' collaborative behaviors over the entire class time. Furthermore, analysis did not show support for significant gain in accuracy when onset features were combined with concurrent features. Finally, we discuss implications for studying collaborative learning and development of software to facilitate collaboration. [For the full proceedings, see ED607784.]