Towards a Multiclass Neural Decision Tree for Large Scale Data Classification

Abstract

Machine learning is an evolving field of artificial intelligence which involves developing algorithms which enable machines to gain insights from data and utilize the knowledge gained for efficient decision making. The field of machine learning integrates several advanced methods ranging from simple linear regression to more complex neural networks. For improved accuracy in binary and multi-class classification tasks, this work proposes the Neural Decision Tree (NDT) classifier which leverages the computational strengths of Artificial neural networks by integrating a neural network at each node of a decision tree. This research is aimed at addressing one of the major shortcomings of decision trees which is the high tendency to overfit which leads to poor generalization to unseen data. The results obtained from the experiments performed using the NDT classifier in this project indicates that the Neural Decision Tree (NDT) performs more efficiently than traditional decision trees in terms of accuracy of predictions and achieves a better generalization to the unseen data. The NDT classifier is used to predict the diabetes status of patients with an accuracy of 97.23% and an AUC score of 0.98, additionally the NDT classifier is applied to the dry bean dataset, achieving an accuracy of 92.77% on the test set. These results demonstrate the versatility of the NDT classifier in various domains including medical applications.