Student Behaviour Modelling and Adaptive Techniques for Social Robots: Data-driven and Teacher-Perceived Evaluations

Abstract User modelling and knowledge representation are important steps towards building personalised systems. Users’ attention and communication are examples of social factors that go beyond simply analysing task efficiency, adding additional complexity to achieving effective human understanding. More specifically, in the educational domain, while the technical performance of adaptive methods plays a primary role in their adoption by researchers, secondary factors, such as teachers’ ability to understand and their intention to adopt, can also influence the implementation and broader acceptance of social robots with adaptive behaviours. In this paper, we validate our high-level proposal for user modelling targeting activities with social robots in the classroom from two different perspectives: the performance of the methods using data from a real-world scenario, and the perceptions of teachers. For the data analysis, various decision-making methods were compared. These included two user-parametrised approaches (a simple rule-based and a fuzzy system, both previously co-designed with teachers) as well as five established supervised machine learning algorithms. For validation of teachers’ perceptions, five teachers were interviewed to gather feedback on their thoughts about our proposal and its practical implications. The findings demonstrate that while teachers initially preferred the semantic modelling offered by the fuzzy system due to its interpretability, three out of five teachers changed their preference after being presented with the results of our data analysis. They favoured the most accurate method over the one they found more intuitive.