Valence-Arousal Modeling

Valence-arousal modeling studies how to represent and predict affect as continuous levels of pleasantness (valence) and activation (arousal), rather than only as discrete emotion labels. Methods include regression, distance-based soft labels and probability distributions, temporal models, and reliability-aware fusion of facial, audio, text, and physiological signals. The approaches support tasks such as in-the-wild affect recognition, tracking emotional trajectories, music recommendation, and controlling emotion in generated text or motion, with benchmarks such as Aff-Wild2 used for audiovisual evaluation. A central challenge is that affect labels and cues vary across people and settings: systems must handle uncertainty and missing or noisy modalities, and it remains unclear when text or other signals help forecast future affect beyond prior trajectory patterns.