Classification is one of the most well-known branches of data mining, which has been widely used in various fields. Today, data classification has a very special place in decision-making, decision-making and diagnosis processes, and it is used in many different industries and sciences. Several different classification approaches have been developed in the literature to increase the accuracy of the results. However, only a few of them can be effectively used to manage collections that contain large sets of data. The main reason is related to the computational time and high cost of such classifiers, which makes them inefficient. Despite their lower accuracy, linear and statistical classifiers are the most popular classifiers for large datasets due to their low computational cost and favorable time. One of the classification methods that has been used in various researches in the subject literature individually or in combination is the logistic regression method for binary data and spaces. In this research, a new learning method is proposed to increase the accuracy of logistic regression classifiers. The method that is used in most researches to train the coefficients of the logistic regression function is the maximum likelihood method, which works based on a continuous objective function. In this research, we present a discrete objective function for training the coefficients in the logistic regression function, contrary to the usual practice. This method uses mixed integer mathematical modeling, and in order to eva‎luate it comprehensively, twenty-three benchmark datasets are used for classification, and its performance is compared with the conventional logistic regression classifier. Experimental results show that the model with the proposed discrete objective function approach achieves a higher classification rate than the conventional logistic regression model in all datasets. In addition, the proposed classifier can achieve an average classification rate of 91.57% on twenty-three benchmark datasets, which is approximately 9% higher than the classical model. The classical model obtained an average classification rate of only 82.50%. Considering this issue, the proposed method can be a suitable and effective alternative to the linear classification approach in dual datasets. This issue is very effective and important especially when we are dealing with big data.

مهدي خاشعي آشياني

علي زينل همداني

نادر شتاب بوشهري , صبا صارمي نيا