FEEL:AfNIRS–EEG Emotion Dataset and Benchmark Library

We present FEEL, a multimodal EEG– fNIRS dataset and benchmark library for emotion recognition. Specifically, the FEEL dataset contains synchronized EEG and fNIRS recordings from 34 participants while they listened to 38 affective music clips, and it also provides rich annotations including discrete emotion labels, dimensional ratings, and liking scores. Based on this dataset, we systematically evaluate representative fusion strategies at the signal-level, feature-level, and decision-level, and we conduct these evaluations using both traditional machine learning methods and deep learning models under subject-dependent, cross-subject, and cross-music evaluation settings. The experimental results show that neurovascular coupling features derived from EEG and fNIRS achieve the highest classification accuracy in subject-dependent evaluations, whereas multimodal fusion approaches do not consistently outperform single-modal methods across all evaluation scenarios, and different fusion strategies exhibit distinct performance characteristics across evaluation scenarios. Finally, we integrate FEEL into the LibEER benchmark framework with extended support for fNIRS and EEG–fNIRS multimodal inputs, providing a standardized benchmark with baseline results for reproducible and systematic multimodal emotion recognition research.