Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Multi-dataset Joint Pre-training of Emotional EEG Enables Generalizable Affective Computing
Authors: Qingzhu Zhang, Jiani Zhong, Zongsheng Li, Xinke Shen, Quanying Liu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 3 Experiments 3.1 Datasets and experiment setup 3.2 Implementation details 3.3 Results of cross-dataset subject-independent classification 3.4 Results of zero-shot generalization to a new dataset 3.5 Impact of more datasets used in pre-training 3.6 Ablation Study Table 3: Performance of cross-dataset subject-independent classification (leave-one-dataset-out evaluation). |
| Researcher Affiliation | Academia | 1Department of Biomedical Engineering, Southern University of Science and Technology, Shenzhen, China 2School of Data Science, University of California, San Diego, La Jolla, CA, USA 3School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China EMAIL, EMAIL EMAIL |
| Pseudocode | No | The paper describes the model architecture and losses in sections 2.1, 2.2, 2.3, and 2.4, and Appendix B, using descriptive text and mathematical formulations. However, there are no clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available at https://github.com/ncclab-sustech/ md JPT_nips2025. and Our code is available at https://anonymous.4open.science/status/ Covariance_EEG_Emotion-D1C1. The code for this research will be open-sourced, and the Git Hub link is provided in the appendix. |
| Open Datasets | Yes | EEG-based emotion recognition is a natural candidate for this paradigm: it has growing availability of labeled datasets (e.g., SEED (7), DEAP (8), FACED (9)) and strong practical value in affective computing and BCI applications. We employed multiple EEG datasets with varying numbers of emotion categories and recording channels for model pre-training (Table 2). We further perform experiments on the Emo EEG-MC dataset (28), a multi-context dataset including imagery-induced paradigm (guided narratives with active imagination), eliciting more internally driven and sustained emotions. |
| Dataset Splits | Yes | We split the target dataset at a 1:3 subject ratio for MLP training and testing. We repeated the random split 6 times and report their mean and standard deviation. A leave-one-dataset-out cross-validation is employed to evaluate the model: the EEG encoder is pretrained on all datasets except the target set. For each testing sample, we compute cosine similarities across all samples in the target dataset and identify its nearest neighbor in the representation space. We evaluate on the imagery task using the few-shot setting: pre-train on six video-induced datasets, then fine-tune on 1/4 of Emo EEG-MC subjects and test on the remaining. |
| Hardware Specification | Yes | All experiments are implemented in Python 3.12.3 using the Py Torch 2.3.1 framework and are executed on an NVIDIA Ge Force RTX 3090 GPU. |
| Software Dependencies | Yes | All experiments are implemented in Python 3.12.3 using the Py Torch 2.3.1 framework and are executed on an NVIDIA Ge Force RTX 3090 GPU. |
| Experiment Setup | Yes | Details of the hyperparameters are shown in Table S1 and Table S2. Table S1: Hyperparameters of pre-training and fine-tuning. Pre-training epochs 20 learning rate 0.0005 weight decay 0.0001 ISA loss temperature 0.07 window length 5 seconds stride 2 seconds Weight of CDA loss 0.02 Fine-tuning batch size 256 learning rate 0.0005 weight decay 0.0022 hidden units 128 epochs 25 |