I am currently working on a binary classification of EEG recordings and I came across with CSP.
As far as I understand, CSP allows you to choose the best features by maximizing the variance between two classes, which is perfect for what I'm doing. Here follow the details and questions:
- I have N trials per subject, from which half belongs to class A and the other half to class B.
- Let's say I want to apply CSP to this subject trials. From what I understood, I should apply CSP to all my trials (please correct me if I'm wrong here). Do I arbitrarily choose which trial from class A to compare with one from class B? Is the order by which I do it, indifferent?
- After CSP I should get the projection matrix (commonly wrote at W), from which I can obtain the transformed signal and compute the variances (part of which will be my features). Why does the computation of the variance is transformed into a log function in most papers?
Thank you very much
There are too many options related to the solutions of CSP in the existing literature, see for example : Chapter A Common Spatial Pattern Approach for Classification of Ment...
Article Decoding Motor Imagery through Common Spatial Pattern Filter...
The projection matrix W is essentially the eigen-decomposition of the covariance of the data matrix X. Alternatively this can also be done using singular value decomposition (SVD) which is a more efficient way of handling high dimensional data as this avoids calculation of the large matrix XT.X. Apply SVD to the entire dataset and then make a plot of the cumulative sum of the SVs against the number of SVs. This can help in selecting the proper number of SVs from the plot which accounts for say 95% of the variances in the data. Log-scale specifies relative changes, while linear-scale specifies absolute changes.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms