A Real-World Data Collection for Emotion Recognition Using Smartwatches and Lessons Learned

Emotion recognition is important across healthcare, education, and daily life, but real-world deployment remains challenging. Collecting reliable biosignal and self-report data in natural settings poses key methodological and design challenges. We introduce "BEmotion", a dataset that links optical and kinematic biosignals to self-reported valence, arousal, and context. The dataset was collected using smartwatches at three job fairs, involving 44 participants. While real-world datasets are essential, improved methods for collecting and using them are needed. Thus, we also analyze the challenges of real-world data collection encountered in the study. From these insights, we derive recommendations as well as lessons learned, and future use cases. We contribute: (1) with an open dataset and (2) recommendations for future dataset collections in the field. Furthermore, we identify a need for greater personalization in emotion recognition rather than a one-fits-all recognition.