Opportunities and Challenges in Generalizable Sensor-Based Affect Recognition for Learning

Recent years have witnessed major research advances in sensorbased affect recognition. Alongside these advances, there are many open questions about how effectively current affective recognition techniques generalize to new populations and domains. We conducted a study of learner affect with a population of cadets from the U.S. Military Academy using a serious game about tactical combat casualty care. Using the study data, we sought to reproduce prior affect recognition findings by inducing models that leveraged posture-based predictor features that had previously been found to predict affect in other populations and learning environments. Our findings suggest that features and techniques, drawn from the literature but adapted to our setting, did not yield comparably effective models of affect recognition. Several of our affect recognition models performed only marginally better than chance, and one model actually performed worse than chance, despite using principled features and methods. We discuss the challenges of devising generalizable models of affect recognition using sensor data, as well as opportunities for improving the accuracy and generalizability of posture-based affect recognition.