Skip to content

Latest commit

 

History

History
31 lines (22 loc) · 1.79 KB

README.md

File metadata and controls

31 lines (22 loc) · 1.79 KB
Raw

EEG Emotion Recognition

Date started: 03.10.2019

Date finished: 09.02.2020

Emotion recognition from EEG data - Graduation thesis

Dataset used: DEAP - 32 people (16 male, 16 female), 40 trials per person, 32 electrodes, 128Hz.

Output classes: For emotion representation the Circumplex Model (arrousal and valence) is employed.

Feature extraction:

  1. Time domain:
  • Statistical: mean of the raw signal over time N, standard deviation of the raw signal, mean of the absolute values of the first differences of the raw signal, mean of the absolute values of the first differences of the normalized signal, mean of the absolute values of the second differences of the raw signal, mean of the absolute values of the second differences of the normalized signal, skewness, kurtosis,
  • Non-linear: mobility, complexity
  1. Frequency domain (FFT):
  • Average and relative power (Welch method) for alpha, beta and theta bands
  • Multi‐Electrode Features: differential and rational assymetry (between left and right hemisphere)
  1. Time-frequency domain (Wavelet transformation)
  • Energy and entropy

Train-test splitting (75 train : 12.5 validation: 12.5 test):

  • Select all the trials for 4 people and keep them as a test set.
  • For the rest of the people (28) and trials do cross-validation.
  • Use GroupKFold (7 groups) because the dataset is ordered. As a result, all of the trials for a specific person are either in the training or the validation set. 1/7 is validation set i.e. all the trials for 4 people (40*4).

Classifiers: Gradient Boosting, AdaBoost, Random Forest, Extreme Gradient Boosting, Nearest Neighbors, Naive Bayes, SVM, Decision Tree. Evaluation metrics: accuracy and F1.

Best classifier: SVM (F1: Valence 73.0, Arousal 67.01).