Although in Chapters 1 and 2 we proposed the strong temporal classification task, the focus in this thesis has been on weak temporal classification. Since we don't know how to do weak temporal classification in a symbolic, comprehensible way, it was felt it was a more fundamental issue.
Extending the metafeatures approach to do strong temporal classification would be worthwhile, but it is probably another thesis of work. However, there are some simple and obvious things that can be done to make inroads into the problem. One potential approach to the problem is a ``weak train, strong test'' approach. This would be based on the ``pre-segmented TC'' approach discussed in Section 2.3.3. TClass would be trained on pre-segmented instances. However, when it came to testing, a ``window'' of the data observed so far could be used to see if it matches any of the learnt classes. We then ``slide'' the window forward and repeat the process to get a sequence of classifications. By using some kind of voting mechanism, it might be possible to actually do strong temporal classification.