A discrete intelligent classification methodology

Abstract

Over the years, classification techniques have been widely used in various fields of application. Intelligent models are among the most popular classification techniques, successfully applied in different science fields. Despite widespread use, intelligent classification models have a fundamental flaw in learning, neglected in the literature. Indeed, the learning process of these existing models is based on a continuous distance-based cost function, which conflicts with the discrete nature of the classification problem. In other words, using this type of function for a classification problem with a discrete objective function is irrational or at least not perfectly rational. The current paper proposes a new classification methodology based on a discrete direction learning-based approach to fill this gap. In order to implement the proposed approach, the multi-layer perceptron model, one of the most famous intelligent models, has been used exemplarily. Although it can be theoretically proven that the performance of the discrete direction learning-based multi-layer perceptron is not worse than its classic version, the proposed model, based on several benchmark datasets from the UCI repository, demonstrates its superiority. The competitive results show that the proposed DIMLP approach achieves a 94.43% classification rate which shows a significant improvement compared to the classic MLP model, which can only reach an 82.13% classification rate. Therefore, the proposed discrete direction learning-based learning approach can be a powerful alternative to traditional intelligent classification approaches. © 2023, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.