Detecting Concept Shifts under Different Levels of Self-awareness on Emotion Labeling

Dataset

• SEED : 3 class (neutral, positive, negative)
• SEED-IV : 4 class (happiness, sadness, fear, neutral)
• GAMEEMO : 9 class (valence, arousal)

Preprocessing (Offline Transform, Feature Extraction ...)

0.0 preprocessing.py
Temporal Segmentation (Sliding Window, Time Partitioning, Time Window, ...)
+ Feature Extraction (Signal Transformation, Dimensionality Reduction, ...)

• Segmentation (Raw Signal)
• Segmentation + DE (Differential Entropy)
• Segmentation + PSD (Power Spectral Density)

  	Seg (channels, window)	Seg + DE (channels, 4 bands)	Seg + PSD (channels, 4 bands)
  SEED	(62, 400)	(62, 4)	(62, 4)
  SEED-IV	(62, 400)	(62, 4)	(62, 4)
  GAMEEMO	(14, 256)	(14, 4)	(14, 4)

EEG channels(num_electrodes), Segment size(Window size)

Make Dataset

train and test data are split for reliable generalization evaluation.
0.1 make dataset.py : ( train : test = 9 : 1 )

---

Our Method

1) Subject-dependent Train, Test

0.2 subdepend.py: For each subject, subdepend.py is executed.

Example.

python ./0.2\ subdepend.py --dataset=GAMEEMO --label=v --model=CCNN --feature=DE --epoch=100

2) Make OOD Detection Model

1.0 OOD detector.py

• Data Pipeline

  • scaling
    • Raw Signal - standardization
    • PSD - log transformation
  • deshape(reshape)
    • CCNN - grid: make 9x9 grid (samples, channels, 4 bands) -> (samples, 4 bands, 9, 9)
    • TSC - expand : (samples, channels, window) -> (samples, 1, channels, window)

• Model

  • CCNN : https://link.springer.com/chapter/10.1007/978-3-030-04239-4_39
  • TSCeption : https://arxiv.org/abs/2104.02935
  • DGCNN : https://ieeexplore.ieee.org/abstract/document/8320798
  • All models are based on TorchEEG : https://github.com/torcheeg/torcheeg

• criterion = torch.nn.CrossEntropyLoss()

• optimizer = torch.optim.Adam(model.parameters())

• scheduler = CosineAnnealingWarmUpRestarts(optimizer)

• Please note the paper to take detailed paremeters.

Example.

python ./1.0\ OOD\ detector.py --dataset=GAMEEMO --label=v --model=CCNN --feature=DE --epoch=100 --cut=6

3) Remove OOD and Train ID

2.0 Base_Remove.py
Set the threshold of ODM to adjust the OOD removal rate (approximately 10%). Please note the paper to take detailed thresholds.

Example.

python ./2.0\ Base_Remove.py --dataset=GAMEEMO --label=v --model=CCNN --feature=DE --epoch=100 --detector=Low_6 --threshold=0.95

4) Compare with the Baseline

In 2.0 Base_Remove.py, threshold==0 is equivalent to Baseline.

Results

---

LICENSE

This repository has a MIT license, as found in the LICENSE file.

Contact

If you have any questions, please email [email protected].