11th Annual COE Graduate Poster Presentation Competition
A neurobehavioral approach to classify trust through machine learning based EEG signals analysis
Student(s): Kazi Farzana Firoz (PhD)
Advisor(s): Dr. Younho Seong
Cross-Disciplinary Research Area: Human Factor, Machine Learning, Neuroscience
Trust is an important factor in any teamwork, including human-human teaming as well as human-machine teaming; therefore, measuring and identifying trust from mistrust of human is important.
- Studies for understanding trust generally done in qualitative method; e.g. questionary, survey
- One of the quantitative ways to measure and calibrate is by observing neural signals.
- Most of the neural studies of trust/mistrust, so far, have been conducted for a specific condition
- Those studies are difficult to interpolate in different context in a generic manner
- Neural phenomena for this construct in a general sense is yet to be understood
- Neural signals can be investigated in time domain and frequency domain.
- In frequency domain, human brain signals can be grouped in several frequency ranges:
-Delta (1-4Hz), Theta (4-7Hz), Alpha (8-13Hz), Beta (15-30), Gamma (30-80)
- Power spectrum analysis in frequency domain can reveal knowledge about human brain function
- Machine Learning techniques are useful for classification analysis
- While studying human brain signals in a laboratory context, real data samples are of scarce
In our study, we aim to:
- Classify trust from mistrust experienced in a ‘basic’ context rather than in a specific context, from neural perspective
- Identify classification approach to achieve the best model to provide the highest accuracy with a small dataset
- We can classify trust/mistrust in a ‘basic’ context wit Machine Learning
- This approach helps us to overcome the limitation to analysis with a small dataset
Oh, Seeung. "An Investigation of Neural Correspondence of Human Trust in Automation." PhD diss., North Carolina Agricultural and Technical State University, 2018.
This research is supported by TECHLAV project sponsored by DoD
Significance of this approach
Original dataset
Technique: repeatedstratifiedKfold cross-validation (n_split=17,n_repeat=10)
Increasing n_repeat to 100 increases the maximum acc to 59.8% with GNB
Nornamlized dataset:
Technique:
repeatedstratifiedKfold cross-validation (n_split=17,n_repeat=100)
Nornamlized dataset:
Technique:
repeatedstratifiedKfold cross-validation (n_split=17,n_repeat=1000)
Data augmentation (TTA, scale=0.02, n_cases=1000)
- This research provides us a path to address the issue of classification of trust/mistrust in a basic context.
- Future research will collect data of trust and using the technique to understand more.
From results, we have observed that
- Performing classification technique directly does not provide a clear idea of which algorithm would be suitable
- For cross-validation with repeatedstratifies K-fold cross-validation, increasing the number of the repeat can improve the accuracy and can provide a better picture of a suitable algorithm
- Normalization preprocessing technique has increased the accuracy significantly
- With data augmentation, we get a robust picture of the accuracy performance of the algorithms as well as provides higher accuracy
- We achieved maximum accuracy of 72% with the Decision Tree Classifier
The dataset contains EEG data from a ‘elicitation study’ (Oh, 2018).
- Words associated with trust and mistrust are shown to particpants (n=17)
- The neural signal is collected as EEG data and
- Power spectrum density of different frequency ranges have been accumulated in this dataset
Analysis:
- Machine learning technique has been used for classification for two classes (trust and mistrust)
- For maximum accuracy, different Machine Learning Algorithms (LR=Logistic Regression, LDA=Linear Discriminant Analysis, KNN=K Nearest Neighbor, DT=Decision Tree, GNB=Gausian Naïve Bayes ) has been tested
- Normalization procedure has been performed
- RepeatedstratifiedKfold cross-validation has been performed
- Data augmentation has been performed
Table 2:results of original dataset
Table 3:results of normalized dataset
Table 4:results of normalized augmented dataset
Table 1:sample rows of collected dataset
Figure 1:Electroencephalography headset with 10-20 system