Skip to content

FEEL Benchmark

Citation: Singh, P., et al. (2026). arXiv:2604.05926.
DOI: 10.48550/arxiv.2604.05926
Access: Paper + code available. Check GitHub for implementation.
Tier: 3

Overview

The first large-scale benchmarking study unifying 19 public datasets for emotion recognition from EDA and PPG signals. Evaluates 16 architectures across 4 paradigms (ML, DL, pretraining, CLSP).

Included Datasets

WESAD, CLAS, DAPPER, K-EmoCon, DEAP, MAUS, CASE, UBFC-Phys, Emognition, PhyMER, Nurse, LAUREATE, ForDigitStress, MAUS, VERBIO, Unobtrusive, MOCAS, Exercise, and more.

Key Findings

  • CLSP (contrastive signal-language pretraining) best for F1
  • Random Forest, LDA, MLP remain competitive (36/114 top scores)
  • Handcrafted features consistently beat raw signals
  • Real-life → lab transfer: F1=0.79
  • Lab → Empatica E4 transfer: F1=0.73

Relevance

Pre-built evaluation pipeline and preprocessing code. Can directly benchmark Meltdown Minder features against all 19 datasets.