MS-MDA

Code of paper MS-MDA: Multisource Marginal Distribution Adaptation for Cross-subject and Cross-session EEG Emotion Recognition

Datasets

The dataset files (SEED and SEED-IV) can be downloaded from the BCMI official website

To facilitate data retrieval, we divided both datasets into three folders according to the sessions, the file structure of the datasets should be like:

eeg_feature_smooth/
    1/
    2/
    3/
ExtractedFeatures/
    1/
    2/
    3/

LOSO Experiments (ATTENTION!)

In our paper, Section III. B. Scenarios mentioned:

we take the first 2 session data from one subject as the source domains for cross-session transfer, and take the first 14 subjects data from one session as the source domains for cross-subject transfer. The results of cross-session scenarios are averaged over 15 subjects, and the results of cross-subject are averaged over 3 sessions. Standard deviations are also calculated.

However, as described in ISSUE 3, LOSO (Leave-one-subject-out) is also required, we therefore additionally evaluated our method in the LOSO paradigm with compared works (In the batch size of {16, 32, 64, 128, 256, 512}). Note that these LOSO experiments are not included in the original paper, and since other works have not yet made their code open-source, we reproduced some of them. The results are shown below (csesn stands for cross-session, csub stands for cross-subject, the number next to it represents batch size, the best result for one transfer scenario is in bold):

Dataset	Method	csesn_512	csub_512	csesn_256	csub_256	csesn_128	csub_128	csesn_64	csub_64	csesn_32	csub_32	csesn_16	csub_16
SEED	DCORAL	65.49±9.95	57.43±8.49	77.33±11.10	62.50±6.77	73.74±9.09	64.05±8.38	74.65±10.46	65.40±9.27	76.24±8.51	64.98±8.42	76.40±10.49	66.39±7.55
	DAN	67.67±8.41	57.11±6.57	79.84±9.42	68.48±6.74	83.15±7.30	71.95±6.55	87.12±7.20	79.03±7.07	88.57±7.60	81.10±6.63	88.88±7.02	82.22±7.09
	DDC	68.32±8.62	57.21±6.52	80.37±8.62	68.98±6.34	84.98±7.64	72.78±8.25	88.16±6.89	78.91±7.62	90.65±6.41	81.27±6.83	89.89±6.69	82.21±7.15
	MS-MDA	79.93±9.90	72.31±10.17	87.68±9.22	78.78±10.70	87.20±10.76	80.33±10.00	89.76±9.03	80.91±9.38	90.38±7.03	81.57±9.81	90.65±8.01	82.67±9.51
SEED-IV	DCORAL	24.80±7.18	20.12±5.31	42.71±9.47	39.77±8.27	48.60±12.53	41.48±7.72	54.39±10.90	45.95±7.19	57.48±11.55	48.12±6.35	59.61±10.03	51.85±7.30
	DAN	27.79±8.43	21.30±4.98	52.62±11.91	41.47±8.02	58.63±11.77	47.50±8.80	67.12±13.47	59.11±7.99	71.51±11.98	63.57±9.07	74.40±12.67	69.68±9.24
	DDC	27.18±6.83	21.94±5.89	53.30±11.42	42.67±9.81	59.00±11.48	48.50±8.44	67.94±12.20	58.18±8.20	71.00±13.25	64.00±8.52	74.81±12.65	69.90±9.95
	MS-MDA	37.75±10.26	36.03±8.99	62.01±13.43	56.20±12.85	64.04±15.27	61.06±12.69	66.87±15.69	62.77±11.23	70.31±15.35	65.12±13.85	72.77±14.71	67.96±11.94

Usage

Run python msmdaer.py, and the results will be printed in the terminal.

Contributing

Issues are welcome. For major changes, please open an issue first to discuss what you would like to change.

Citation

If you find our work useful for your research, please consider citing our paper as:

@article{chen2021ms,
  title={MS-MDA: Multisource Marginal Distribution Adaptation for Cross-subject and Cross-session EEG Emotion Recognition},
  author={Chen, Hao and Jin, Ming and Li, Zhunan and Fan, Cunhang and Li, Jinpeng and He, Huiguang},
  journal={arXiv preprint arXiv:2107.07740},
  year={2021}
}

@inproceedings{chen2021meernet,
  title={MEERNet: Multi-source EEG-based Emotion Recognition Network for Generalization Across Subjects and Sessions},
  author={Chen, Hao and Li, Zhunan and Jin, Ming and Li, Jinpeng},
  booktitle={2021 43rd Annual International Conference of the IEEE Engineering in Medicine \& Biology Society (EMBC)},
  pages={6094--6097},
  year={2021},
  organization={IEEE}
}

TODO

• LOSO experiments on SEED and SEED-IV, methods including DDC, DAN, DCORAL, MS-MDA, on two transfer scenarios (cross-subject, cross-session)
• ISSUE 6

License

This source code is licensed under the MIT license