data set IIb
Matthias Kaper, Peter Meinicke, Ulf Grossekathoefer, Thomas Lingner
University of Bielefeld
For classification, we used an SVM-Algorithm. After Bandpass-filtering
(0.5-30 Hz) and scaling, we trained the SVM using data from a subset of
electrodes with a timeframe of 0-600ms poststimulus. For this, we used an
equal number of samples for positive and negative examples (negative randomly
chosen). Of course, best parameters were chosen using crossvalidation on a
grid of the both (hyper-) parameters bandwidth and C.
For testset classification, we used the best parameters and classified each
timeseries within a block of 12 using the same subset of electrodes.
Afterwards, the margin value was summed up for corresponding columns/rows
over different repetitions. The highest values should then correspond to the
"correct" column and row.
It is worth to mention, that we took the series, the way they were. For the
first "repetition", we only used the first 12-block, disregarding any other
block. This seemed to be the most realistic approach to us (probably not the
one with the highest classification values). Similar, the second repetition
applied only the 12-blocks 1 and 2. Long story short: We kept the
chronological sequence.
Actually, the strategy was almost the same as in our nips-2002-paper
("Improving transfer rates..."). We tried several other configurations but
achieved best perfomance with this approach.