Social robots are expected to be a human labor support technology, and one application of them is an advertising medium in public spaces. When social robots provide information, such as recommended shops, adaptive communication according to the user's state is desired. User engagement, which is also defined as the level of interest in the robot, is likely to play an important role in adaptive communication. Therefore, in this paper, we propose a new framework to estimate user engagement. The proposed method focuses on four unsolved open problems: multi-party interactions, process of state change in engagement, difficulty in annotating engagement, and interaction dataset in the real world. The accuracy of the proposed method for estimating engagement was evaluated using interaction duration. The results show that the interaction duration can be accurately estimated by considering the influence of the behaviors of other people; this also implies that the proposed model accurately estimates the level of engagement during interaction with the robot.