Short-Term Emotion Change Prediction With Compact and Informative Mobile Sensing Inputs

Abstract

Emotional change prediction can help characterize affective dynamics from real-world behavioral data and may inform future non-clinical affective-state estimation systems. This study investigates short-term emotion change prediction using only objective mobile sensing features. Based on the StudentLife dataset, emotion change is defined as the difference in Photographic Affect Meter (PAM) scores over a three-day window, without introducing subjective questionnaire variables as model inputs. A modeling pipeline incorporating lag features, anomaly-related processing, and ensemble regression is adopted, and systematic comparisons are conducted across different sensor modalities and hourly time-window configurations to evaluate predictive performance under low-burden input settings. Results show that, with full inputs, the model achieves an MSE of 2.08, an Accuracy@2.0 of 88.62%, and an extreme-sample error of 0.21. Under a reduced-input setting, using only the unlock sensor with the hr17 time window still achieves an MSE of 4.38, an Accuracy@2.0 of 79.69%, and an extreme-sample error of 0.83. These results indicate that predictive information for short-term emotional changes is not evenly distributed across all modalities and time periods; instead, a small set of key inputs can preserve behavior signals of practical value. Further interpretability analysis shows that duration-related features, including duration_per_unlock and unlock_duration, contribute more strongly than simple frequency-based features within the unlock modality. The findings suggest that identifying high-information features for specific populations may inform the design of lower-burden affective-state estimation systems, while reducing redundant data collection and potential privacy exposure.