Extraction of Classification Rules from Enhanced Fuzzy Min-Max Neural Network

Abstract

Though neural networks have capability in solving complicated problems, their deficiency of being ‘black box models’ has prevented them from being acknowledged as a typical practice for applications such as real-time applications, restorative medical and business for predictions. Enhanced fuzzy min max neural network (EFMN) has a number of enhancements to the original fuzzy min max neural network (FMN). Hence it is more accurate than original FMN and other related classification algorithms. Like other artificial neural networks (ANNs), EFMN is also like a black box and the knowledge is expressed in terms of min-max values of the hyperboxes and associated class labels. So the justification of classification results given by EFMN is required to be obtained to make it more adaptive to the real world applications. This paper proposes a model to extract classification rules in the form of if then from trained EFMN using partial decision trees (PART) algorithm. These rules justify the classification decision given by EFMN without loss of the accuracy. For this, EFMN is trained for the appropriate value of θ. The min-max values of all the hyperboxes and their respective class labels are given as input to PART algorithm which extracts rules by repeatedly generating partial decision trees from the input instances. These rules are readable and represents the trained network. These rules give accurate justification of the classification decision. The applicability of the proposed method is tested on widely used Fisher’s Iris dataset.