Constructing transformation rules for inductive learning
Evolving Transformation System (ETS) [Goldfarb, 1992] captures inductively the compositional structures of a class through an evolving set of distance functions applying on the data attribute values. It composes of a set of transformation rules, which transform one datum to another without loss of class information. By capturing the object relationship within class, the class structure can then be modeled. The transformation rules differ from logical rules in describing classes in that the relationship between objects within class is modeled, whereas logical rules only model objects within class independently. Thus an Evolving Transformation System describes a class's composition while the dissimilarity between classes can also be known. This thesis investigates the constructions of the transformation rules for class characterization. New design and applications in data with different pre-defined structural information are experimented in the construction of the transformation rules. The learning algorithm is applied to molecular sequence data and a number of different bench-marked data sets for inductive learning. The data sets include: (1) the parity-3 data, (2) a set of cancerous and healthy cells, (3) a set of blood pressure data and (4) a set of machine learning data.