Classification is one of the most important branches of data mining, which has been successfully used in a wide range of problems. In the literature, many efforts have been made to provide more efficient and accurate classification models. The presented classification models can be classified into different categories from the perspective of different dimensions, the most important of which are the categories, their classification into various statistical, intelligent and deep models. Despite the wide variety of classification models presented in the literature, all of them have been developed based on a similar methodology, and a fundamental issue related to the learning process of these models has been ignored. In fact, in the learning process of all existing classification models in the literature, a distance-based continuous cost function is used to estimate the unknown parameters of the model, even though the classification problem has a discrete nature. In other words, in the usual learning processes where the training of the classification model is based on distance-based continuous cost functions, the fitting values continuously approach the real value. In other words, in each iteration of the training, the distance between the actual and fitted values decrease continuously. Therefore, in continuous modeling problems such as causal predictions or continuous time series analysis, it can generally be guaranteed that the performance of the model is continuously improved. However, this issue is not true in discrete modeling such as classification problems. In fact, the final outputs a continuous learning process need to be discrete before reporting as the output of the classification model. As a result, if the learning and discretization processes are aligned, the learning process may lead to improved performance. Otherwise, the effect of the learning process will be lost under the influence of discretization. Therefore, the learning process in such conditions can be useless and ineffective. Thus, using a continuous cost function for the classification problem, which itself has a discrete objective function, can affect the accuracy of the results. On the other hand, using a learning process based on discrete functions can lead to better results due to its greater compatibility with the discrete nature of the classification problem. In this thesis, a different classification methodology that uses a discrete function in the learning process has been proposed and implemented. For this purpose, three logistic regression classification models, multilayer perceptron neural network and deep multilayer perceptron neural network, which are among the most popular statistical, intelligent and deep classification models respectively, are used as examples to implement the proposed methodology. In order to comprehensively eva‎luate the proposed classification approach, the proposed models and their classic version have been implemented on several benchmark datasets and their performance has been compared. The results of this thesis indicate the superiority of the classification models based on discrete direction-based learning compared to their classical versions. Therefore, the proposed classification approach based on discrete learning can be used as a useful alternative method in learning processes of different classification models.

مهدي خاشعي آشياني

علي زينل همداني

عليرضا بصيري , صبا صارمي نيا , اميرحسين اميري