Metafeatures are a novel feature construction technique that can be applied whenever there is some kind of underlying substructure to the training instances and there is some way to extract these substructures. In temporal domains, these substructures take the form of sub-events, like intervals of increase or decrease.
Metafeatures first extract substructures within the training instances. Then, interesting, typical or distinctive examples are selected. These substructures become synthetic features, which are then fed to a propositional learner. We can then convert the output of the learner back to a human-readable form that is described in terms of the metafeatures we began with.
The extraction of typical or distinctive examples is done by segmenting the parameter space into regions. The division of the parameter space into regions can be approached using two methods: traditional undirected segmentation which simply treats each instance as a point in space; or the novel directed segmentation approach. The directed segmentation approach is specifically designed to facilitate subsequent learning by dividing the space into regions where the class distribution is non-uniform; hence trying to find regions which could potentially be useful for discrimination between classes.
The novelty of metafeatures lies not in the idea of parametrised substructures within training instances; these are ideas that have been used for ad hoc temporal classification for a very long time. However, the novelty lies in these metafeatures as a general concept, and that they form a parameter space. The novelty lies in segmenting the parameter space in such a way as to:
Finally, two novel enhancements are suggested: