How can we integrate all of these data sources? Since all of these processes produce attribute values, they can simply be concatenated into a single long attribute value tuple. This leads to the diagram shown in Figure 5.6.
Note that there are several new ``boxes''. The first of these is the learner; and the second is attribute combination. The attribute combiner takes attributes from multiple sources and concatenates the attributes from the different sources into a single attribute value tuple, ready for passing to a propositional learner.
Certain aspects of TClass are not included in Figure 5.6, in particular the creation of human-readable output. In Section 4.10.2, we discussed post-processing the learnt concept to generate more human-readable descriptions in situations where the generated classifier represents the concept as a set of constraints on the attribute values. This requires several sources: firstly and most obviously, the learnt concept; secondly, the synthetic features to post-process the concept description, and thirdly, the original instantiated features to generate the bounds on values. The practical implementation of this is discussed in Section 5.5.7.
The architecture for testing is almost the same as that for training, and is shown in Figure 5.7. The most notable difference is that the synthetic features generated by parameter space segmentation in the training stage are reused.
