Clustering Technique for Evaluating and Validating Neural Network Performance

Abstract

Data used for training and testing a neural network (NN) are often collected from limited sample projects. They may constitute clusters instead of being evenly distributed over the entire space. This paper first studies the effect of clustered data on the performance of an NN model by fitting a cowboy hat surface, followed by an introduction to the fuzzy clustering technique. An NN model is then evaluated cluster by cluster over a representative space. New predictions are validated based on their locations in the space and the model performance in corresponding regions. The analysis improves the confidence of a user on an NN model.