The PEPs presented are simple and do not identify domain-specific features. Surprisingly, they perform relatively well. It is possible improve performance by designing multi-channel PEPs (for example, rather than treat x, y and z separately, a PEP could treat them as a single unit). Furthermore, the PEP design did make use of a useful property of temporal domains: correlated events on different channels.
The classifiers used here do not use feature subset selection at all. Previous research [John et al., 1994] shows that learners may benefit significantly from feature selection, especially in cases like this one where so many features are generated by event clustering.
The descriptions produced could be significantly improved over those generated by C4.5, since C4.5 is designed to tell classes apart; as a result, it does not always produce characteristic descriptions, rather it prefers discriminant ones [Quinlan, 1998]. While this is appropriate for a tree, it is not necessarily appropriate for building intuitive descriptions of temporal classes. Future work will focus on alternative learning algorithms for producing comprehensible descriptions.