An investigation of TREPAN utilising a continuous oracle model
William A. Young suffix II suffix, Gary R. Weckman, Murtaza Rangwala, Harry S. Whiting suffix II suffix, Helmut W. Paschold, Andrew H. Snow, Chad Mourning · International Journal of Data Analysis Techniques and Strategies · 2011
TREPAN is decision tree algorithm that utilises artificial neural networks (ANNs) in order to improve partitioning conditions when sample data is sparse. When sample sizes are limited during the tree-induction process, TREPAN relies on an ANN oracle in order to create artificial sample instances. The original TREPAN implementation was limited to ANNs that were designed to be classification models. In other words, TREPAN was incapable of building decision trees from ANN models that were continuous in nature. Thus, the objective of this research was to modify the original implementation of TREPAN in order to develop and test decision trees derived from continuous-based ANN models. Though the modification were minor, they are significant because it provides researchers and practitioners an additional strategy to extract knowledge from a trained ANN regardless of its design. This research also explores how TEPAN’s adjustable settings influence predictive performances based on a dataset’s complexity and size.