Rule extraction from a trained neural network using principle interval projection with certainty factor

Abstract

Mining data with a neural network is one of the most significant techniques. But understanding how each feature implies the results is not simple due to the difficult interpretation of those mathematical operations of the network. Hence, extracting meaningful and understandable knowledge in a simple conditional English sentence from a trained neural network is essential. This dissertation proposes a technique for extracting knowledge based on the activation interval projection on each dimensional axis with certainty factor refinement. Without any mathematical expressions, the knowledge is captured in forms of if-then rules that the premises are the conjunction of input feature intervals representing in linguistic quantities. The experimental results from the rule extraction process show comparable classification accuracy with neural networks and can be used with the real world applications such as Wisconsin breast cancer database.