رشته های مرتبط مدیریت
گرایش های مرتبط بازاریابی
مجله محاسبات عصبی – Neuro computing
دانشگاه گروه مهندسی صنایع، Özyeğin، استانبول، ترکیه
نشریه نشریه الزویر
1. Introduction Classification algorithms label the observations of a data set using their different attributes. In recent years, Artificial Neural Network (ANN) classifier has become popular because of its broad application areas. In most of these applications there is a focus on cost-sensitive learning as there are different costs for different types of misclassifications [1–6]. Cost of misclassification of an instance is different from one context to another. In most of the cost-sensitive binary classifications such as diagnosis problems, there are two different misclassifications (false negatives and false positives) and each of them has a particular cost. However, in business problems such as credit card fraud detection and direct marketing each misclassified observation can have a different cost and moreover there may be a profit for correctly classified ones (true positives and true negatives). Thus, in such cases, there is a necessity to develop a classification model that considers individual profits and costs. “Credit card (CC) fraud detection” is one of the well-known classification problems. In this type of classification, a data set which contains various information (attributes) about the credit card transactions of a bank is used [7–10]. For each of the records, there is a dependent attribute which takes the value of one if the transaction is fraudulent and takes the value of zero if it is legitimate. In CC fraud detection if a classifier correctly detects a fraudulent transaction, it will save the usable limit of the corresponding card and if it mis-classifies the transaction, the usable limit of the card will be lost. The other application of data mining and classification used in this study is Bank direct marketing. In order to target a specific segment of customers, banks use data mining algorithms to classify the customers as buyers and nonbuyers [11–13]. In this context, if a model correctly detects a potential customer for a campaign, there will be a particular profit of gaining that customer and if a potential buyer is not identified, the profits that could be gained from him/her might be lost. These two applications of classification are discussed in this study, where individual costs and profits of each instance is considered. Although ANN has become a popular classification algorithm in recent years, there are a few studies about variable individual costs of misclassifications. Excluding this issue and using the simplistic model, ANN will just minimize the number of misclassifications which will result in a suboptimal solution in terms of cost minimization or profit maximization . In real life problems, the most important reason which motivates business administrations to invest in data science is the amount of profit they gain from implementing their models. In credit card frauds, a possible fraud transaction with a high amount of available limit is more important for bank to be detected than a transaction with a low usable limit. Equivalently, in the direct marketing context (for a bank), a customer with a high potential balance is more important to be detected and receive an offer than a customer with a lower balance.