Abstract
The development of a top-down, hierarchical modeling architecture called a "knowledge tree" is introduced. Such trees comprise of nodes with local models that approximate data at the specific nodes. By clustering, a tree is unfolded successively. The novelty of this approach lies in the ability to refine the clusters and models in a step-wise fashion compared to single global rule-based structures such as the Takagi-Sugeno model. More so, local models at each node augment the overall architecture with high functionality compared to a decision tree whose nodes compute some broad classification rules. Experimentally, using synthetic and real-world data sets, compared to a single global model, a hierarchical tree is built faster and more efficiently while retaining high accuracy. Being hierarchical and step-wise, knowledge trees can be promising in large-scale applications (data of high dimensionality).
