BCI COMPETITION 2005
DATASET V, preprocessed features
Contributor: Irene Sturm (student submission)
Co-Contributor: Guido Dornhege
Fraunhofer FIRST (IDA), Berlin, Germany
Classification was done with Regularized Discriminant Analysis
( Reference: J.H. Friedman, Regularized Discriminant Analysis, Journal
of the American Statistical Association, vol.84(405), 1989.).
The parameters were chosen with crossvalidation. For each subject the chosen
value of the parameter lambda was 1, which means that the resulting
classifier is linear (Regularized Linear Discriminant Analysis, RLDA).
Beforehand a channel selection based on training a Linear Programming Machine
was performed. The selected channels were
- for subject 1: C3, Cz, C4, CP1 and CP2.
- for subject 2: C3, Cz, C4 and CP1.
- for subject 3: C3, Cz, C4, CP1, CP2 and Pz.
One sample in the feature vector is the estimated PSD over the last second,
there are 16 samples per second. The mean of the classification output of 8
consecutive datapoints was computed in order to provide an output every 0.5
seconds. This means that the first label in the result file uses the first 8
given samples, the second one uses samples 9-16 and so on...
The labels were provided in an ASCII-file in continuous order in one column.
- for subject 1 in subject1.txt
- for subject 2 in subject2.txt
- for subject 3 in subject3.txt
Note that the performance can be enhanced enormously by using more than one
second of data which can be done, e.g., by applying a moving average of 3 or
4 seconds to the classifier output.