Conclusions

TClass has proved to be an effective learner in terms of accuracy if one is willing to use a voting mechanism. In every domain, voting different learners allowed it to exceed the performance of our baseline learners. Furthermore, on the two real-world domains, it proved competitive even with hand-selected domain-specific features. This is a somewhat surprising result.

However, clearly voting leads to a destruction of the readability of the classifier. This is a definite problem with the current implementation of TClass. However, the definitions it produces, while not as accurate as those created by voting, are still very informative. In the artificial domains, we saw that the induced concepts corresponded closely to the correct concepts. Similarly, in the real-world domains, correlations between known concept descriptions and the induced concept descriptions are visible. However, as noise is increased, the readability of the induced rules is reduced.

There is a question as to which kinds of domains and problems TClass is suited to. Results show that TClass performs better, even without voting, in the following situations:

• Where there is little noise in the signal. This was particularly revealed in the TTest artificial domain, in Figure 6.18. This diagram showns that TClass' performance is much better than the baseline learners at low noise levels, but as the noise level becomes very great, the baseline learners like Naive Segmentation and Hidden Markov Models perform better.
• Where there are fewer training instances. This was particularly shown in 6.20,, where TClass outperformed other classifiers when there were fewer training instances.