Trustable Decision Tree Model using Else-Tree Classifier

TLDR: In this article , the Else-Tree classifier is proposed, which allows the classification model to learn its limitations by rejecting the decision on cases likely yield to misclassifications and hence produce highly confident outputs.

Abstract: Abstract With advances in machine learning and artificial intelligence, learning models have been used in many decision-making and classification applications. The nature of critical applications, which require a high level of trust in the prediction results, has motivated researchers to study classification algorithms that would minimize misclassification errors. In our study, we have developed the {\em trustable machine learning methodology} that allows the classification model to learn its limitations by rejecting the decision on cases likely yield to misclassificationsand hence produce highly confident outputs. This paper presents our trustable decision tree model through the development of the {\em Else-Tree} classifier algorithm. In contrast to the traditional decision tree models, which use a measurement of impurity to build the tree and decide class labels based on the majority of data samples at the leaf nodes, Else-Tree analyzes homogeneous regions of training data with similar attribute values and the same class label. After identifying the longest or most populated contiguous range per class, a decision node is created for that class, and the rest of the ranges are fed into the else branch to continue building the tree model. The Else-Tree model does not necessarily assign a class for conflicting or doubtful samples. Instead, it has an else-leaf node, led by the last else branch, to determine rejected or undecided data. The Else-Tree classifier has been evaluated and compared with other models through multiple datasets. The results show that Else-Tree can minimize the rate of misclassification.