1) Baseline: Initially we considered the easiest/most
favourable scenario, where we have a balanced stream and
every label is available. In this scenario, all except the
FixedUncertainty Strategy achieve 100% accuracy and Kappa
statistic after processing the 1k instances. This shows that with
enough labelled data is possible to learn an accurate classifier
for PD detection. The low performance of the FixedUncertainty
derives from the default threshold value of 0.9 and
shows the a big limitation of this strategy since in a on-line
setting is very challenging to determine the threshold value
that leads to good results.