The CBF dataset turns out to be largely trivial as a learning problem. All three approaches, TClass, naive segmentation and hidden Markov models attain 100 per cent accuracy on it. It may be possible to prove theoretically that a naive segmentation approach with 20 segments can be guaranteed with some very high probability to produce a learner which can classify instances in the CBF domain perfectly. The effect of smoothing on the outcome appears to be minor in this case; but slightly positive overall for the TClass classifiers.
This makes the CBF domain a poor one for evaluation of some of the other issues we are interested in; since it is likely that in many cases we hit the ``100 per cent barrier''; which does not give good guidance to the real-world performance of the classifier.
However, one thing that can be compared as a result of these experiments is comprehensibility; in particular between naive segmentation and TClass. While both results are readable and provide some useful description of the learnt concept; we claim subjectively that the TClass definitions are easier to understand, as they are expressed in terms that a human looking at the data might use: the height of maxima, gentle changes in gradient, and sudden changes in gradient; rather than averages of particular temporal segments of the data.