Results (
Thai) 1:
[Copy]Copied!
While there is much redundancy between the measurements
(for instance the feature pairs b/d and f/j above), it was
decided to use all the possibilities since the SVM is not much
affected by this and performs well under situations of correlation
between data. In fact, in a later stage, we empirically
noted that removing some of the features, which seemed not
to be part of individual filling style characteristics, gave
improved results. Ultimately, employing some feature decorrelation
(or selection) procedure might be helpful. To
perform this, a much larger experiment would be necessary
for allocating part of the data set for development.
Being translated, please wait..
