Collaborative game-based learning environments have significant potential for creating effective and engaging group learning experiences. These environments offer rich interactions between small groups of students by embedding collaborative problem solving within immersive virtual worlds. Students often share information, ask questions, negotiate, and construct explanations between themselves towards solving a common goal. However, students sometimes disengage from the learning activities, and due to the nature of collaboration, their disengagement can propagate and negatively impact others within the group. From a teacher's perspective, it can be challenging to identify disengaged students within different groups in a classroom as they need to spend a significant amount of time orchestrating the classroom. Prior work has explored automated frameworks for identifying behavioral disengagement. However, most prior work relies on a single modality for identifying disengagement. In this work, we investigate the effects of using multiple modalities to detect disengagement behaviors of students in a collaborative game-based learning environment. For that, we utilized facial video recordings and group chat messages of 26 middle school students while they were interacting with Crystal Island: EcoJourneys, a game-based learning environment for ecosystem science. Our study shows that the predictive accuracy of a unimodal model heavily relies on the modality of the ground truth, whereas multimodal models surpass the unimodal models, trading resources for accuracy. Our findings can benefit future researchers in designing behavioral engagement detection frameworks for assisting teachers in using collaborative game-based learning within their classrooms.

Recent years have seen growing interest in modeling visitor engagement in museums with multimodal learning analytics. In parallel, there has also been growing concern about issues of fairness and encoded bias in machine learning models. In this paper, we investigate bias detection and mitigation techniques to address issues of algorithmic fairness in multimodal models of museum visitor visual attention. We employ slicing analysis using the Absolute Between-ROC Area (ABROCA) statistic to detect encoded bias present in multimodal models of visitor visual attention trained with facial expression and posture data from visitor interactions with a game-based museum exhibit about environmental sustainability. We investigate instances of gender bias that arise between different combinations of modalities across several machine learning techniques. We also measure the effectiveness of two different debiasing strategies—learned fair representations and reweighing—when applied to the trained multimodal visitor attention models. Results indicate that patterns of bias can arise across different modality combinations for the different visitor visual attention models, and there is often an inherent tradeoff between predictive accuracy and ABROCA. Analyses suggest that debiasing strategies tend to be more effective on multimodal models of visitor visual attention than their unimodal counterparts