Architecture

TClass implements the architecture suggested in Figures 5.1 and 5.2. As a historical artifact, it also implements several other architectures (for example, the per-class clustering approach discussed in [Kad99], and also covered in Section A.2). It also implements several baseline algorithms (for example, the naive segmentation approach, discussed in Section 6.1.1). There are classes that correspond directly to the combiners, feature extractors and learners discussed in Section 5.1. In particular, three things are implemented using inheritance; making it very easy to implement additional components:

• Global feature extractors: These are objects that extract aggregate global features, such as means, and global maxima and minima. Conventional attribute values (like patient age and gender) are easily implemented within this framework.

• Metafeatures: Metafeature objects encapsulate the extraction functions that are applied to training streams, as well as the characteristics of that metafeature's parameter space.

• Parameter Space Segmenters: Segementers segment the parameter space into regions marked by synthetic events. As discussed in Section 4.8, this includes undirected segmenters like K-Means and directed segmenters like the random segmenter.

• Learners: These are conventional attribute value learners. This is a thin wrapper around the functionality provided by Weka.

Each of these components can be mixed and matched. For example, any metafeature can be used with any segmenter.

For added flexibility, each of these components has a mechanism for setting whatever settings are useful to it. For example, if one is applying K-Means, one may wish to specify the number of clusters or ask that it be determined automatically.