Estimating and improving students' engagement in a collaborative learning environment is an important component in the field of learning research. Collaborative learning is a strategy of learning activities employed by small groups in which cooperative learning behaviors are closely related to other members or objects in the group. Researchers showed that students who are actively involved in class learn more. Therefore, gaze behavior and facial expression are important nonverbal indicators in cooperative learning environments. In this paper, we proposed a multimodal deep neural network (MDNN) to solve the engagement prediction problem in collaborative learning. We combined facial expression and gaze direction as individual streams of MDNN to predict engagement levels in collaborative learning environments. Our multi-modal solution was evaluated in a real collaborative environment with a significant accuracy of 74%. The results show that the model can accurately predict students' performance in the collaborative learning environment.