Comparison of Resampling Techniques for Covariance Matrices Derived From Small Datasets With Shrinkage Regularization for Event-Related Potential Data
Keywords
Loading...
Authors
Issue Date
2021-06-18
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Since electroencephalogram data in the eld of brain-computer interfacing
often has a poor signal-to-noise ratio, covariance matrix estimations for
methods such as linear discriminant analysis (LDA) can become imprecise.
Therefore it can be bene cial to improve these covariance matrix estimations
such that further processing of the data becomes more accurate. This
can be done with a number of techniques like Ledoit-Wolf shrinkage regularization.
I propose in this thesis that covariance matrices might also be
improved by means of resampling the data to nd better estimations. I compare
one resampled LDA pipeline implemented with di erent hyper parameter
settings against a simple shrinkage regularized LDA pipeline. However,
running these pipelines on two event-related potential datasets revealed that
the resampling techniques do not signi cantly improve the performance of
the LDA over the simple shrinkage regularized LDA and that compared to
the shrinkage regularized LDA the resampled LDA has a worsened run time.
In the large dataset, the performance of the resampled LDA became significantly
worse from the shrinkage regularized LDA. Given these results, the
resampling techniques thus do not improve the covariance in the way it was
implemented in this thesis.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen