Feel Emotion Eval

feel-emotion-eval

FEEL: Quantifying Heterogeneity in Physiological Signals for Generalizable Emotion Recognition — Singh et al. (2026) (arXiv:2604.05926, 2026)

What this evaluates

Evaluates the generalization and transferability of emotion recognition models across heterogeneous physiological signal datasets. It probes how well different modeling paradigms (handcrafted features, raw signal deep learning, and contrastive pretraining) perform under subject-independent, cross-dataset, and low-data regimes.

Datasets

• WESAD — total ?; splits: LOSO-CV (-1)
• NURSE — total ?; splits: LOSO-CV (-1)
• EMOGNITION — total ?; splits: LOSO-CV (-1)
• UBFC_PHYS — total ?; splits: LOSO-CV (-1)
• PhyMER — total ?; splits: LOSO-CV (-1)
• EmoWear — total ?; splits: LOSO-CV (-1)
• MAUS — total ?; splits: LOSO-CV (-1)
• CLAS — total ?; splits: LOSO-CV (-1)
• CASE — total ?; splits: LOSO-CV (-1)
• Unobtrusive — total ?; splits: LOSO-CV (-1)
• CEAP-360VR — total ?; splits: LOSO-CV (-1)
• ScientISST MOVE — total ?; splits: LOSO-CV (-1)
• Dapper — total ?; splits: LOSO-CV (-1)
• ForDigitStress — total ?; splits: LOSO-CV (-1)
• ADARP — total ?; splits: LOSO-CV (-1)
• Exercise — total ?; splits: LOSO-CV (-1)
• MOCAS — total ?; splits: LOSO-CV (-1)
• LAUREATE — total ?; splits: LOSO-CV (-1)
• VERBIO — total ?; splits: LOSO-CV (-1)

Metrics

• F1 scores (primary) — range: [0, 1]
  • Average F1 score computed across all Leave-One-Subject-Out (LOSO) cross-validation folds or participant-wise train/test splits. For multi-class tasks, macro-averaged F1 is standard in this domain.
• accuracy — range: [0, 1]
  • Average classification accuracy computed across all LOSO folds or train/test splits.

Input / output format

Input: Time-series physiological signals (EDA, PPG, or both) sampled at device-specific rates. For raw-signal deep learning models, inputs are segmented into 60-sample windows with 50% overlap.

Output: Discrete emotion labels: binary (high/low arousal or valence) or four-class (combinations of arousal/valence quadrants).

Scoring recipe

def compute_metrics(predictions, gold_labels, folds):
    fold_f1s = []
    fold_accs = []
    for train_idx, test_idx in folds:
        y_true = gold_labels[test_idx]
        y_pred = predictions[test_idx]
        fold_f1s.append(f1_score(y_true, y_pred, average='macro'))
        fold_accs.append(accuracy_score(y_true, y_pred))
    return {'f1': np.mean(fold_f1s), 'accuracy': np.mean(fold_accs)}

Common pitfalls

• Class imbalance mitigation differs by paradigm: SMOTE/oversampling is applied to ML/DL models but explicitly avoided for CLSP fine-tuning except in severe imbalance cases.
• Evaluation splits are strictly subject-independent (LOSO-CV) or participant-wise (50/50 train/test), so subject leakage must be avoided.
• Cross-dataset generalization uses Leave-One-Dataset-Out (LODO) and zero-shot baselines, which require careful domain alignment and cannot be directly compared to in-domain LOSO results.

Evidence (verbatim from paper)

Performance was measured using average accuracy and F1 scores, standard metrics in physiological emotion recognition (Schmidt et al. (2018); Singh et al. (2024)) computed across all LOSO folds.

Citation

@misc{singh2026feel,
  title={FEEL: Quantifying Heterogeneity in Physiological Signals for Generalizable Emotion Recognition},
  author={Singh et al. (2026)},
  year={2026},
  note={arXiv:2604.05926}
}

• arXiv: 2604.05926