Binary hierarchical multiclass classifier for uncertain numerical features

Real-world multiclass classification problems involve moderately high dimensional inputs with a large number of class labels. As well, for most real-world applications, uncertainty has to be handled carefully, unless the classification results could be inaccurate or even incorrect. In this paper, we investigate a binary hierarchical partitioning of the output space in an uncertain framework to overcome these limitations and yield better solutions. Uncertainty is modeled within the quantitative possibility theory framework. Experimentations on real ultrasonic dataset show good performances of the proposed multiclass classifier. An accuracy rate of 93% has been achieved.

I. Introduction

Multiclass classification represents a substantial task in data mining and machine learning applications. It is a typical technique working under supervised learning mode. It consists, in the first time, in inducing a classifier from labeled training data with pre-defined classes and afterward, involving the induced classifier to classify new unlabeled data. Multiclass classification is commonly used in numerous real-world applications including security [1], natural disaster forecasting [2], psychology [3], medicine [4], etc.