A Review of Fraud Detection Algorithms for Electronic Payment Card Transactions
الموضوعات :Touraj BaniRostam 1 , Hamid BaniRostam 2 , Mir Mohsen Pedram 3 , Amir Masoud Rahamni 4
1 - IAUCTB
2 - IAUCTB
3 - Associate Professor of Electrical and Computer Engineering Department, Faculty of Engineering, Kharazmi University
4 - SRBIAU
الکلمات المفتاحية: Unsupervised Algorithm, Fraud Transaction, Accuracy, E-Card, Supervised Algorithm,
ملخص المقالة :
Abstract—Several studies have been presented to solve challenges of electronic card (e-card) fraud that the two main purposes of these studies are to identify types of e-card fraud and to investigate the methods used in bank fraud detection. To achieve this purpose, one of the most common methods of detecting fraud is to investigate suspicious changes in user behavior. Supervised learning techniques help to find anomalies by analyzing user behavioral history based on past transaction patterns in fraud detection systems. One of challenging issues in detecting fraud is to consider the change of customer behavior and the ability of fraudsters to devise new patterns of fraud, which makes unsupervised learning techniques popular for detecting unknown and new frauds. In this paper, the concepts of fraud, types of banking fraud along with their challenges, different form of fraud and banks' data research tools for early identification have been examined, then a review of the researches on fraud detection will be conducted. This paper aims to introduce fraud detection techniques and methods that have provided appropriate results in the big data environment. Finally, the fraud detection algorithms and proposed methods of related works presented in this paper, will be fully compared on a common dataset in terms of parameters such as speed of fraud detection, accuracy, and cost (hardware and network resources). Ensemble Meta-Learning can be used alone to build a stronger classifier. These techniques have been relatively successful in detecting fraud and reducing costs.
10
Journal of Advances in Computer Engineering and Technology
A Review of Fraud Detection Algorithms for Electronic Payment Card Transactions
Received (Day Month Year)
Revised (Day Month Year)
Accepted (Day Month Year)
Abstract—Several studies have been presented to solve challenges of electronic card (e-card) fraud that the two main purposes of these studies are to identify types of e-card fraud and to investigate the methods used in bank fraud detection. To achieve this purpose, one of the most common methods of detecting fraud is to investigate suspicious changes in user behavior. Supervised learning techniques help to find anomalies by analyzing user behavioral history based on past transaction patterns in fraud detection systems. One of challenging issues in detecting fraud is to consider the change of customer behavior and the ability of fraudsters to devise new patterns of fraud, which makes unsupervised learning techniques popular for detecting unknown and new frauds. In this paper, the concepts of fraud, types of banking fraud along with their challenges, different form of fraud and banks' data research tools for early identification have been examined, then a review of the researches on fraud detection will be conducted. This paper aims to introduce fraud detection techniques and methods that have provided appropriate results in the big data environment. Finally, the fraud detection algorithms and proposed methods of related works presented in this paper, will be fully compared on a common dataset in terms of parameters such as speed of fraud detection, accuracy, and cost (hardware and network resources). Ensemble Meta-Learning can be used alone to build a stronger classifier. These techniques have been relatively successful in detecting fraud and reducing costs.
Index Terms—Accuracy, E-Card, Fraud Transaction,
Supervised Algorithm, Unsupervised Algorithm.
2 Assistant Professor of Computer Engineering Department., Islamic Azad University, Central Tehran Branch, Tehran, Iran. (banirostam@iauctb.ac.ir).
3 Associate Professor of Electrical and Computer Engineering Department, Faculty of Engineering, Kharazmi University, Tehran, Iran. (pedram@khu.ac.ir).
4 Full Professor of Computer Engineering Department, Science and Research Branch, Islamic Azad University, Tehran, Iran. (rahmani@srbiau.ac.ir).
I. INTRODUCTION
I
llegal use of card information without the owner's knowledge is a fraudulent act [1]. Fraud involves any deliberate or premeditated act to deprive or misuse money or any other personal property through criminal act, deception or any other unfair practice [2]. In order to prevent the imposition of damages caused by fraud, various activities have
been carried out under the titles of inhibition, prevention and detection of fraud. Fraud detection actually involves monitoring the behavior of millions of users to estimate, detect, or prevent undesirable behavior. If it is possible to predict and identify the behavior of cardholders on the line and with high accuracy and prevent their fraudulent behaviors, it will be easy to attract customers' trust and increase their satisfaction.
According to McAfee, failure to detect fraudulent transactions will impose huge financial losses on enterprises, which will reach about $44 billion by 2025 [3]. Surveys show that although U.S. credit card transactions have the highest rate, but it has the lowest rate of fraud [4]. But at the top of the list were Ukraine with an astonishing rate of 19%, Indonesia with 18.3%, Yugoslavia 17.8%, Turkey 9% and Malaysia 5.9% the highest rate of fraudulent transactions [5]. The above reports show the extent and importance of the problem of electronic banking fraud. Therefore, detection of fraud has become one of the most important research topics.
Because of the unbalanced nature of bank transaction data, it is very difficult to identify whether transactions have been authorized or fraudulent. Various methods are used to detect e-card fraud, most of which are data mining-based [6]-[8].
Data mining methods are one of the main tools for detecting fraud in e-cards. If the data volume and the amount of difficult relationships between them increase, the access to the information embedded in the data will be more difficult. Thus, the role of data research as one method of knowledge discovery in identifying the allowed user will be more important [106].
Allowed users using some evaluation factors such as credit card number, cardholder's address, signatures and expiration date may allowed to make credit card transactions [9], [10]. Allowed users usually display certain behavioral attribute. Therefore, fraud detection models based on a set of patterns, including information about the usual consumption category, the last time of transaction, the amount of credit consumed, evaluate the amount of change in customers' behavior. In this regard, Users behavioral patterns and machine learning algorithms are used to analyze and identify anomalies [11], [12].
Machine learning algorithms have a variety of parameters that are much more efficient at identifying user behavior patterns than those not identified by an expert system.
The development of common machine learning algorithms has led them to solve certain problems. One of the most important is that normal data distribution is balanced [13], [14] but this factor does not practically apply in banking transactions and distribution of fraudulent and non-fraudulent transactions is not balanced.
Because only a tiny number of cases are fraudulent, a model that accurately detects 99% of allowed transactions, despite its high accuracy, does not help us find fraud items with high accuracy [15]. In such data, we may experience the phenomenon of over-fitting. This causes the accuracy of the model to be reduced in predicting data that does not exist in sampling and mislead the model.
Following in the section II, the subject literature on the frauds and challenges of fraud detection, how to aggregate information, method of computational fraud detection systems, reliable fraud detection systems, and finally the most widely used algorithms for detecting e-card fraud will be investigated. In the section III, 14 solutions and related researches will be introduced, in the section IV, comparison of different parameters of widely used algorithms will be presented and finally in the section V the conclusion will be presented.
II. Subject Literature
This study aims to investigate a model for detecting fraud in businesses that use big data. In particular, the main purpose is to detect fraud patterns in e-card transactions and focus on organizing and analyzing the phases of big data.The methods for dealing with fraud in payment systems are divided into two categories: anomaly prevention and anomaly detection under Figure 1.–
Fig. 1: Fraud in payment systems.
A- Anomaly Prevention Method
In the anomaly prevention method, when a transaction is run, the transaction is compared to the sample of previous patterns and previously known attacks, and if the similarity is detected, the transaction is identified as an attack.
B- Anomaly Detection Method
In the anomaly identification method, it is attempted to create a feature of the performance history for each user and the user's profile is formed, after extracting any deviation large enough (depending on the accuracy of the algorithm) it is recorded in the user's set of attributes to identify and warn the user in case of any attack [16].
A secure and reliable banking payment system requires simultaneously labeling suspicious transactions besides requiring legal user authentication [17]-[20]. Basically, fraud prevention and identification are both important parts of risk management in e-cards. It is virtually impossible to decide easily on the purpose of doing a transaction legally or illegally. The best and least costly idea is to track fraud using mathematical methods for data [21]-[23]. Of course, there are naturally two main criticisms of data mining-based fraud detection research: lack of real data available for research and lack of appropriate published research methods [24], [25].
Regardless of whether the fraud detection system is considered manually or systemically, it should detect fraud correctly and have a low error percentage. It also checks the transaction situation in real-time and discovers the user's ambiguous behavior before completing the transaction [26]-[28]. Transferring data processing after storage to pre-storage operations will significantly reduce the time to assess new demands from the system, and make a detailed decision to detect fraud [29]-[31].
Types of E- Card Transactions
According to Figure 2, e-card transactions can be divided into two categories [32].
Fig. 2: E-card Classification.
A- Physical
In the physical way, the cardholder physically presents his card during the purchase in order to commit fraud; the thief must also remove the card [33]-[35].
B- Virtual
In virtual mode, only some important card information such as card number, expiration date, security code is stolen. These purchases are usually made on the Internet or on the phone, and most of the time the original cardholder is not aware of the stolen information [36], [40], [104]. According to card usage habit analysis, Fraud detection is a promising solution [41], [42].
E-Card Frauds
According to Figure 3, e-card frauds are divided into seven categories:
Fig. 3: E-Card Frauds.
A- Direct(Card-not-present) Fraud in E-cards
This category of fraud is divided into two categories: offline and online. In offline fraud, physical applications are the main frauds through the use, manipulation and copying of cards. In the online fraud section, the use of telecommunication and communication tools is commonly used, and fraud is usually carried out without the presence of the card owner. In Card-not-present, the most of fraud is carried out using card information [44]-[45].
B-Telecom Fraud
This means using communication and telecommunication services to achieve other types of crime and fraud. Individuals, businesses and telecom service providers are among the main victims of this category of fraud [46], [48].
C- Intrusion into the Computer
Intrusion is the act of entry without prior notice and permission. This means potential attempts at unauthorized access to information and their targeted manipulation. The infiltrator may belong to any environment, i.e., an external hacker or an intra-organizational influencer [49].
D- Fraud with invalid Cards
This means using an e-card for which there is no creditor or maintaining account. This method is one of the most complex methods of fraud for prediction [50].
Theft Fraud/ Counterfeit Fraud
Theft refers to the use of a card whose owner is not present, and usually the card is blocked as soon as the customer declares it to the card service provider. But counterfeit occurs when only card information, such as its number and password, is used [51]-[53].
F- Application Fraud
This type of fraud occurs when the requester submits incorrect specifications for the e-card. There are two modes to find this type of fraud. First, the requests belong to a user who is commonly referred to as a duplicate card; second the requests are from different users, but with a same profile, known as fraudulent and identity forgers [54].
G- Behavioral Fraud
This type of fraud occurs when a fraudulent act is carried out using a normal card and in the owner's presence.
Challenges of Detecting Fraud in E-Cards
According to Figure 4, the challenges of detecting fraud in e-cards were presented in nine categories [55]-[57].
Fig. 4: Challenges of Detecting Fraud in E-Cards.
A- Unavailable Real Datasets
One of the most important limitations in this area is the lack of datasets on which researchers can develop their systems [58].
B- Asymmetric Datasets (unbalanced)
Datasets of e-card transactions strongly have Skewness. That means the ratio of fraud transactions to legal transaction is very low, which makes it difficult to explore the pattern. Usually, in real terms, 98% of data is legal and only 2% of them are fraud. When using skewed data, many learning algorithms perform poorly [59]-[61].
C- Size of Datasets
Millions of e-card transactions are made every day. Analyzing such huge amounts of information requires highly effective and scalable computational methods that will also require considerable resources. This matter generally imposes many limitations on researchers [62]-[64].
D- Dynamic Fraudulent Behavior
Fraudsters have dynamic behavior, i.e. over time they change their behavior against the system. Therefore, frauds are becoming more complicated over time, as far as even the experts of this era cannot predict them. Thus, this issue has turned the detection of fraud into a decision in terms of ambiguity [65]. Almost all fraudulent transactions have a legal appearance, and if each of them is examined separately, there is no doubtful sign in them, while these transactions increase the possibility of detecting doubtful items when checked simultaneously along with other transactions related to the card [66]-[69]. In parallel with the growth of fraud prevention technologies, fraudsters also show more advanced behavior, in addition, fraud events are hidden among the bulk of legal data, and by changing fraudulent behavior, the classification of fraud will be inefficient, detection of the fraud or legality of the incoming transaction is not a reliable decision [70]-[72].
E- Dispersion of Fraud Events
The most important challenge fraud detection systems are facing with is to mis-classify legal transactions as fraud [73]. This leads time-wasting and improper use of organizational resources, as well as customer uncertainty, and besides e-cards, the dispersion of fraud events causes uncertainty in the accuracy of data which usually leads to false reporting of fraud [74]-[76].
F- Independent Parameters of Detection Pattern
When implementation a detection model, it is very important to use those features that lead to a correct classification. When implementation a fraud detection model, the initial set of features (raw features) contains information about a single transaction value. Many developed fraud detection models have used raw transaction features such as time, value, location and other transaction features. Of course, the use of these raw transaction data will not lead to the extraction of customer's behavior and cannot be considered as a standard method for detecting fraud patterns [77]-[79].
G- Model stability
Fraud patterns change. Rule-based systems are not efficient against the management of emerging patterns.
H- Difficulty of the Matching Patterns
Some fake patterns are very similar to normal and legal patterns. Detecting these types of frauds will be very complicated [80]-[83].
I- High Dimensions
For process, the number of transactions to detect fraud is usually too large. Some process millions of transactions.
Aggregation of Transaction Information
The important point when aggregating transaction information is interval is appropriate to go back from previous transactions, because the longer the time goes back, the less reliable recent transaction information receives. For this purpose, some researchers suggest time intervals of 24, 60 or 168 hours. Some also suggested that only transactions be processed from the S transaction during the tp hours before the i transaction[84], [85]. In their researches, some suggest investigating transactions within the same interval. These features include: the number of transactions made during D hours, the financial volume carried out during D hours, and finally, the number of terminals used during D hours. The only point remains is to calculate the time parameters for a period like the above. Using statistical indicators such as averages can be completely misleading. For instance, the average transaction hours for a card that has a transaction at the beginning of the morning or around noon, which does not show the actual habit of the card owner[86]-[89]. Therefore, other approaches should be concentrated for this purpose. Crabensen et al. suggest using Vonmises distribution. This distribution is also known as periodic normal distribution, which is a distribution of a normal distribution variable enclosed in a circle.
This distribution is defined as Equation (1) for a basis such as D = {t1,…,tn}, all of which are radians[90]:
Computational Fraud Detection Methods
Computational fraud detection methods are divided into two subcategories of supervised and unsupervised [91]. One question of researchers has always been, which approach is more appropriate? Dal Pazolo et al. have sought to answer questions such as what machine learning algorithm should be used by examining previous work on detecting e-card fraud. Are static models sufficient, or are dynamic models needed to be felt?
Static models are updated in long-term intervals, but dynamic models are being updated at any moment. Now this question arises, how many transactions are enough to train a model? Should data be checked in its skewed format or balanced by re-sampling methods? What performance indicator is appropriate? Finding a specific value is actually a data mining method that is widely used to detect fraud [92]. Artificial defense systems, auditing, artificial intelligence, database, distributed and parallel calculations, econometrics, expert systems, fuzzy logic, genetic algorithm, machine learning[107], neural networks, pattern exploration, statistics, visualization, etc. have been used. There are several specialized solutions for detecting fraud in businesses such as e-cards, e-commerce, insurance, retail, telecommunications industries. An interesting idea that has been a loan shift from spams is to understand the time nature of fraud in blacklists by following the frequency of words, and classified them in the features of fraudulent observations [93].
Reliable Fraud Detection System
Under Figure 5, a reliable fraud detection system includes the following five properties:
Fig. 5: Feature of reliable fraud detection system.
A- Skewness Distribution
Because of the small number of fraudulent transactions, a fraud detection system should be able to manage and distribute data. One strategy to overcome this challenge is to divide train data into different sections [94].
B- Noise
A dataset can contain incorrect data. A reliable fraud detection system should be able to control this data because noise will affect classification performance.
C- Data overlap
A system that is suitable, which is capable of detecting fraudulent transactions similar to non-fraudulent transactions.
D- Dynamics
The techniques used by fraudulent change. Therefore, a suitable fraud detection system should be able to adapt to new fake patterns [95].
E- Appropriate Classification Criteria
The parameters for evaluating implementation results should be proportional to the dataset, target field and data distribution. For example, the Accuracy criterion is not suitable for skewed data.
Fraud Detection Algorithms
Figure 6 introduces nine widely used algorithms in fraud detection, each of which will be compared separately to speed fraud, accuracy and cost on a shared dataset.
Fig. 6: Fraud Detection Algorithms.
A- Artificial Neural Networks
Artificial neural networks are one of the most well-known methods of fraud detection [37], [38]. This method is used both as supervised and unsupervised. Back propagation of error as supervised algorithms and Self Organization Map (SOM) neural networks as an unsupervised algorithm are used [39].
B- Bayesian Networks
In Bayesian networks, the purpose is to correctly predict the value of a discrete class of variables in Bayesian rules by using the vector of predictors or features [96]. The fraud network is built using experts' knowledge, while the network of ordinary and legitimate users is built using the behavior of non-fraudulent users [16]. Basically, Bayesian networks are faster than neural networks, but this superiority exists until they encounter new observations, because they will act slower [97].
C- Support Vector Machine
It is a supervised machine learning method used to find anomalies. Generalization and time efficiency have shown better, but comparing its performance with neural networks after error propagation in detection of fraud in card transactions has shown that when the number of data is low, the backup machining performs better than neural networks after the error is published, but when faced with larger data sets, then error propagation is more appropriate [43].
D- K -Nearest Neighbor (KNN)
With e-cards, this method also needs to define a criterion for measuring distance in order to evaluate two data instances, in KNN calculations; each incoming transaction is classified by calculating the nearest point. If the nearest point has fraud, the transaction would represent as fraudulent. The performance of this algorithm can be improved by using a genetic algorithm to calculate the distance index. It's a quick technique but comes with a wrong alert [105].
E- Decision Tree
It is one of the powerful algorithms for categorization and detection. The tree uses information theory and entropy to select the best variable and start scrolling. In the fraud detection system, effective variables that play an important role in customer identification in the banking system will be extracted and the role and coefficient of effect of each one will be determined in the transaction database. Then, common decision tree algorithms such as Chaid, EX_Chaid, C4.5 and C5.0 can be used by evaluating the accuracy criteria to identify suspicious behaviors.
F- Systems based on Fuzzy Logic
In 2002, fuzzy neural networks were used to detect fraud in card transactions for the first time to allow more customers to be processed in parallel. Different fields such as the value of transaction, the time interval between transactions, transaction duration, transaction code and duration of sending in transactions are considered as input parameters. In another study, Darwinian fuzzy methods have been used to detect fraud in e-cards. This method has a very high accuracy rate, low, wrong alert, but high time cost [47], [86].
G- Artificial Immune Systems
Artificial immune systems are created by imitating the operation of the immune system. Artificial immune systems in practice are a supervised clustering method that has shown considerable success in classification issues, especially in fraud detection. With detecting fraud in e-cards, insiders are a finite set of legal transactions, and outsiders represent all transactions that are not found in the history of insider ones [98]-[101].
H- Genetic Algorithm
This algorithm is an optimization and searching method that seeks to find at least one approximate optimal solution. In data mining, this algorithm, especially in selecting variables, in combination with other data mining methods, is used to improve SVM parameters to predict the amount of fake, to combine with neural networks for detecting fraud of e-cards at high accuracy rates, as well as artificial immune systems to reduce the wrong alerts in fraud detection [102]. Genetic algorithm is also used to detect error transactions in e-cards to reduce the number of wrong selected transactions [103].
I- Hidden Markov model
Hidden Markov model is a development of the Markov model, in which there is a possible function of modes. The model results from a dual random process, one of which is not visible, but its image can be seen through another set of random processes that produce sequences of observations. With fraud detection, this model is usually one of the unsupervised models that detects distance-based anomalies with performance sequences [46].
III. Related Works
In the following, fourteen related works have been introduced and reviewed. Related works were presented on a common dataset with Kaggle name in part 4 in Table 1 in terms of speed of fraud detection, accuracy and cost (hardware and network resources) will be compared with different fraud detection algorithms. Kaggle dataset contains 492 fraudulent transactions with 284,807 transactions in which it is severely unbalanced. (0.172% is fraudulent). This Dataset, in terms of data confidentiality, has not made main information more available [58]. E-card fraud detection algorithms introduced and related tasks have already been done on dataset based on the parameters implementation or sometimes have been compared by us to other related tasks and presented algorithms are simulated and implemented.
1- Zamini, M et al. in [42] compared the power of impact of prediction algorithms in three supervised classification models and examined logistic regression, sloping reinforced trees and deep learning and It also explores the advantages of creating features using domain allocation and measurable features using Autoencoders and deals with an efficient method. It creates six different feature sets from the main variables. This research compares the features of the above models and concludes that by creating features using domain allocation, a significant improvement in predicting power is made. In addition, Autoencoder provides a way to reduce the dimensions of data and increase the predicting power.
2- Misraa, S et al. in [50]presented a fraud detection model that used Autoencoder and classification algorithms to extract basic features from input data. Their model detects fraud on credit card transactions. A transaction includes many features, including the time and amount of the transaction, the type of transaction (deposit or withdrawal), the customer's account number, their age, the location of the ATM used, etc.
Unnecessary features may lead to poor performance of classification algorithms and ultimately to poor performance of the model, and as a result, the model will be costly in high dimensions in terms of time complexity. Therefore, their proposed model focuses on predetermined features and by examining the correlation coefficient of the features, only selects the significant features and reduces the dimensions. To select the feature, the score got from cross information from a feature and class, statistical filter method is used. But it should be noted that complex relationships that multiple features may have with each other, have been ignored. On large datasets wrapper-based feature selection techniques are computationally costly. Also, none of wrapper methods can guarantee optimal results.
To solve these problems, Autoencoders is presented that effectively handles the reduction of dimensions of input data by discovering the correlation of features. The proposed Autoencoders architecture comprises Encoder to map input data to the hidden layer and Decoder which operations data conversion received from the hidden layer and delivered to the output layer is responsible. The input layer comprises six neurons, the hidden layer of three neurons, and the output layer of six neurons. The activation function is also nonlinear.
In the first step of the proposed architecture, Autoencoder is trained using transaction features. Therefore, Autoencoder encrypts the input data features and provides it to the hidden layer. Considering that the data dimensions in this layer are reduced, in the second step, learning classification will be easier. In this step, the classification will be taught with the labeled data in the previous step. The presented architecture is a general model and in the second step, depending on the user's requirements, any classification can be used. In this paper, three different classifiers, including Multi-Layer Perceptron, K-Nearest Neighbor and Logistic Regression have been implemented and evaluated on the data.
3- Carcilli, F et al. in [30] have used the Multilayer Feed Forward Neural Network model to detect bank transaction fraud quickly. The main question of their research was how to identify fraudulent transactions using neural networks and considering previously known frauds. Therefore, their approach has been to focus on card transactions. They have used real data, but due to lack of labels, they have simulated doubtful transactions by experts' knowledge and subject literature. The neural network used by them has 15 independent input variables, which according to the values they obtained, in the first layer is 61 neurons, and finally its morphology is with the sigmoid tangent transfer function. According to the model, they reported favorable performance in terms of four indicators of correct notification rate, incorrect non-notification rate, Geometrics mean, and Fisher beta statistics.
4- Suman Mishra, J. et al. in [54] have tested a back propagation of error, simple Bayesian and C4-5 as the base classifier on separate data sections, resulting from placement sampling. They used a meta-classifier to select the main class and combined the extracted classifications to create a basic prediction based on concurrent learning algorithms in the construction of a low-cost model for detecting vehicle insurance fraud.
5- F.Carcilli, F. et al. in [20] presented an approach to managing complex problems, combining classification models, and building a meta-classifier to extract different behaviors and achieve a general classification model. This approach has presented a regulatory approach based on the mixture of experts, negative correlation learning and subspace projection method for boosting combination. The artificial intelligence engine, the main classification used in this method, is the ensemble neural network. The method of mixture of experts allows each class to focus on a subset of the train space and explore it well. In addition, a dynamic network estimates weights for the final combination based on the efficiency of each expert (classification). Negative correlation learning also offers a rule for training, an increase in the model's generalization. Finally, the subspace projection method is an approach to increase the train space that can be considered differently. This method has compensated for separate defects of each method for the composition of the classifications. The presented approach is presented for multi-objective regression classification problems that firstly have a large number of features, secondly they have complexity and ultimately have few observations.
6- Randhawa, K. et al. in [3] presented a real-time data processing architecture to detect fraud using spark tools and streaming data. In their architectural design, it is used to manage the queue of messages, data and process the high volume of data in real time. Their proposed architecture includes three key modules for real-time behavior anomaly discovery through machine learning. Logstash data collection module is the next module related to message queue, in which Kafka actually acts as a data passage for streaming data storage and finally the Spark class module for machine learning is real-time. Logistic algorithm implemented regression using Spark ML library on the proposed model. Because in dataset the ratio of transactions with non-fraudulent data labels to transactions with fraudulent labels is high, to achieve greater efficiency, they considered a threshold value in their model and examined the values lower than the base. According to the presented results, it was observed that the best performance was when they considered the threshold value to be 0.6, which is the highest value under the curve equal to 0.92. The proposed model for real-time processing of big data, the possibility of fast, scalable processing and fault tolerance, has provided a good performance.
7- TranP, H. et al. in [57] use the Self Organization Map network to create a framework for detecting fraud. Using visualization of user behavior and an LU decomposition matrix, it has set a threshold for measuring abnormal behavior. And this method has been tested in three problems: detecting telecom fraud, e-cards and intrusion into computer networks. This approach benefits from this property, which is unsupervised and does not require labeled data. Because the database is usually very bulky, there are very few labeled fraud samples. Therefore, the use of supervised approaches has many limitations. Thus, the approach of detecting fraud used in this work is the discovery of pert points. In addition, because of the increase in simplicity, effectiveness and achieving high computational efficiency, the problem is defined as a simple bipolar classification. This means that the neurons with the most members are in the center, are considered as the main neurons, and with one radius, their surroundings will be as an outlier point.
8- Sathyapriya, M. et al. in [58] use random processes and Hidden Markov Model to present a model for detecting fraud that does not require labeled data and apply the card holding habit to detect fraud in e-cards. In this method, the sequence of bank card transactions is extracted by Hidden Markov random processes. Since the small items purchased by the card owner in a transaction are not usually visible to the card issuance bank, using a clustering method, transactions were placed in three clusters only in terms of financial amount. The change in each cluster was an observational sequence. Meanwhile, the cluster with the highest frequency of transaction membership was considered as a profile and played an essential role in the initial evaluation process for training the Markov Model. In each transaction, random processes are built based on the amount of money spent. This method has reported a low, wrong notification rate. To select the parameters N, R and δ, one is changed over time and the rest is kept constant. The correct and incorrect alert value is checked. The number of hidden stated cases was changed from 5 to 10. Sequence length was considered from 5 to 25, and the threshold limit was changed from 0.1 to 0.7. Finally, sequence length or R was 15, threshold 0.3 and number of 10 stated.
9- Song, H. et al. [47] focused on discovering abnormal values in large distributed data environments. The method they use is to discover density-driven anomalies, in which for each row such as p, the factor calculates the local outlier factor value or LOF. This factor shows the extent to which p is likely to be abnormal or outlier relative to its neighboring points. LOF represents the position of the p detachment according to the neighbors, which includes it. In order to implement their model in a distributed big data environment, they use a grid-based segmentation algorithm to preprocess data. The algorithm first breaks all the information space into multiple grids or grids, then assigns each grid to computational nodes in a distributed environment. With this approach, the workload is balanced on each computational node and the flow of the information network decreases. At the time of calculation, they also proposed a distributed system for calculating LOF as DLC1, which aims to find the anomalies in the data in parallel and has two aspects: First, based on LOF features, rows are classified in a two-level grating, i.e. local grid and passing grid. Local grid rows are processed locally and the network will only be required to process passing grid. Then an optimized method is presented to reduce the number of rows that need to be flown into the network. Through a set of simulation experiences, they evaluated and showed that their density-driven anomaly detection method has showed good performance.
10- Gyamfi, N. et al. in [43] use the combination of SOM neural network and the support vector clustering model have presented a model for discovering online influence in computer networks. In the first proposed model, transactions are clustered by a SOM neural network and a set of winning neurons is formed. The SOM neural network can summarize a space with many features to a two-dimensional space. In the next step, the normal and non-normal boundaries of the resulting set are determined using the clustering of the support vector. The labeling of two sub-spaces takes place in the same section. The first input transaction is measured by the SOM winning neurons and assigned to the nearest one, then labeled based on the subspace to which the neuron belongs based on the clustering of the support vector. The resulting model is simulated on a dataset and two standard datasets are measured by adaptive hierarchical SOM method and SOM combination with K-Means, which has shown better results in terms of algorithm performance and run time in classification.
11- Jain,Y. et al. in [16] using the SOM model, they created a normal behavior profile for the card owner. They first stored the weight matrix as profile using previous transactions and the SOM and then measured the distance between the input transaction and the resulting profile and examined the normal or abnormal amount. In addition, with a developmental process, a profile was also composed of detected fraud transactions and the new transaction distance was compared to it. Therefore, two thresholds, one to measure the extent of legality and one for the amount of fraudsters, are defined. The data used were got by simulation. The lack of real environment data in this research has questioned the effectiveness of this research, but low, wrong alert rate in simulated data has
shown promising results.
12- Shpyrko, V. et al. in [61] implemented a set of algorithms on payment transactions to investigate and identify fraudulent transactions from non-fraudulent ones. The proposed algorithms were implemented in real time and their main purpose is to use different machine learning algorithms to achieve higher accuracy with maximum validation score. Therefore, the main problem is the implementation of a model that can detect fraud in real time and block fraudulent transactions to provide a better and safer experience for the user.
All Algorithms such as, Logistic Regression, Support Vectors Method (SVM), K-Nearest Neighbors, Decision Tree Classifier and Artificial Neural Networks were applied on the dataset. The dataset in question is extremely unbalanced, they concluded unbalanced data had problems with over-fitting and incorrect correlation. Over-fitting leads to the identification and prediction of all transactions in a non-fraudulent manner. In fact, incorrect correlation will cause a lack of understanding of the target function, so unbalanced data will lead to an ambiguous correlation matrix. Dataset must be converted to balanced mode through data balancing techniques (random sampling), i.e. the ratio of fraudulent and non-fraudulent data should be 50 to 50. Modeling large volumes of data and high-speed calculations requires powerful computing machines.
There are various data processing tools, but the implementation and modeling used in this research were Pandas Python libraries for simpler data processing, Matplotlib for visualization, Numpy and SciPy for scientific calculations, Seaborn for statistical data visualization, and also Sklearn and Tensorflow were machine learning library.
13- Nath Dornadula, V. et al. in [62] presented a new method of detecting fraud for streaming data of transactions with a target field, to analyze the details of customers' past transactions, extract their behavioral patterns and counter the Concept of Data Drift. This concept is actually a variable that changes over time in unforeseen ways. The existence of such variables will lead to high imbalances.
The main purpose of this paper is to overcome the concept of Data Drift for real-world implementation and functionality. At first, cardholders were divided into different groups (high, medium and low) by clustering based on transaction value. Then, using the Sliding-Window method, it collects some transaction features (such as the maximum and minimum transaction value and the average amount in and time elapsed) to find their behavioral patterns from groups.
Then, an algorithm whose inputs are cardholder ID, t (transaction order sequence) and w )window size( is presented to collect the features. The output of the algorithm is the details of the collected transaction and the actual features of the cardholders or fraud. When a new transaction enters the window, the old transactions are deleted and then processed in the second phase of the collected transactions. Then, the second algorithm, whose inputs are the previous and current transaction ID and its output, is the model score after each transaction, to update the classification score to achieve the accuracy of the model. It should be noted that PCA technique has been used to convert and reduce features to maintain the confidentiality of dataset transactions. SMOTE technique has been used to overcome the unbalanced dataset problem. But Over-sampling hasn't delivered good results.
14- Thennakoon, A. et al. in [63] deals with four cases of fraud that have occurred in the real world. Each of them has been evaluated using examples of machine learning algorithms. This evaluation has provided a comprehensive guide for selecting the optimal algorithm according to the type of frauds and evaluating the performance. Another factor considered in this article is real-time fraud detection. The dataset in question is a combination of fraudulent transaction logs and all transaction logs.
The fraudulent transaction log file keeps all online credit card frauds observed while the log files of all relevant bank transactions are stored within a specified period. Because of retention, confidentiality is sensitive features such as encrypted card numbers. The number of fraud records is 200, while the log file is 917781 records. Therefore, dataset data is extremely skewed and unbalanced. They first divided the raw data into four categories according to the fraud pattern and the information got from the bank.
These four types include Merchant Category Code (MCC), transactions over $100, ISO Response risk transactions, and transactions with unknown web addresses.
These four categories are divided into two different ways: 1. is to convert raw data - numerically, 2. use non-numerical data without conversion - that category 1 of categories 1, 2 and 3 is applied as numerical conversion and 4 is applied non-numerically. Data cleanup preprocessing operations, data aggregation, dimension reduction, data conversion on numeric data, data clearing operations and data aggregation have been applied. They also overcame on unbalanced data.
SMOTE, CNN and RUS methods have been used to balance the data. The 10-fold technique was also used. Modeling was applied to the dataset by four machine learning algorithms: SVM, Naive Bayes, KNN and Logistic Regression. The results show that the use of sampling has provided higher efficiency.
IV. Comparison of different parameters of previous algorithms
Considering to the above-mentioned works, it should be noted that supervised methods, such as neural networks and Baysian networks perform significantly better on labeled data than unsupervised methods, such as Gaussian hybrid models, for detecting anomalies [65], [66]. In short, however, the problem with most of these methods is that they require labeled data, both legal transactions and fraud transactions, to train the classification model [67], [68]. Access to fraud data is one of the biggest problems ahead of fraud detection. This approach cannot detect new frauds for labels that have not existed before. Furthermore, there are the following criticisms regarding any method that uses supervised methods [69]. These criticisms include:
A- In an event-based operating environment, they have a high computational cost.
B- The time required to label new observations is the same as the time required to label previous observations. The learning curve of these methods has a zero degree slope.
C- Deviations due to sample selection can cause errors in train data labels.
D- In environments where observations are personal information of individuals, employees' access to these labeled observations violates privacy policies.
Regarding algorithms, it should be noted that the complexity of the problem of detecting fraud in e-card transactions has caused the use of an algorithm alone, not to cover all aspects of the problem and subsequently provide acceptable accuracy. Algorithms, regardless of supervised or unsupervised, have been compared with each other in terms of fraud detection speed, accuracy and cost of train, the result of which is listed in Table 1. It should be noted that in the previous methods, some parameters of Table 1 were marked with the sign "---", meaning that the proposed method of related work in its evaluations did not focus on the desired parameter and evaluated the other parameter.
[1] Distributed LOF Computing
TABLE I
Comparison of evaluation criteria, type of algorithm, implementation framework and cost of different fraud detection algorithms.
Method | Big Data | Evaluation Criteria | Algorithms | Cost ( Hard ware resources and network) |
Zamini, M et al. [42] | --- | Accuracy=0.989, AUC=0.961 | Proposed Autoencoder | Expensive |
[50] Misraa, S et al. | --- | Accuracy=0.9994, Precision=0.8534 Recall= 0.8750, F1-Score= 0.826 | Autoencoder with a Single Encoding Layer and a Single Decoding Layer. | Expensive |
Carcilli, F. et al. [30] | Kafka, Spark, Cassandra | Precision=4.3 | Distributed Implementation of a Balanced Random Forest, Streaming Procedure | --- |
Suman Mishra, J. et al [54] | --- | Accuracy=0.82 | Hidden Markov Model | Cheap |
F.Carcilli, F. et al. [20] | --- | Per=0.71 | Transaction-Based Case Study | --- |
Randhawa, K. et al. [3] | --- | Accuracy=0.9993, ACC=0.942 | AdaBoost and Majority Voting Methods | --- |
TranP, H. et al. [57] | --- | (OCSVM):Accuracy= 0.9660 Precision=1, FPR= 0.0850 F-Score=1, DR(Recall)=1 (T 2 Chart): Accuracy= 0.9360 Precision=1, FPR= 0.1600 F-Score=1, DR(Recall)=1 | Two Real Time Data-Driven Approaches Using Optimal Anomaly Detection Techniques for Credit Card Fraud Detection (OCSVM, T 2 Chart) | --- |
Sathyapriya, M. et al [58] | Spark | Accuracy=0.82 | Hidden Markov Model | --- |
Song, H. et al. [47] | --- | AUC=0.97(k=30) | Hybrid Semi-Supervised Anomaly Detection Model(A Deep Autoencoder + Ensemble - Nearest Neighbor Graphs-) | --- |
Gyamfi, N. et al. [43] | Spark | 0.84 | Support Vector Machines with Spark | --- |
Jain,Y. et al. [16] | --- | Accuracy: SVM:94.65%, ANN:99.71% BayesianNetwork:97.52% KNN:97.15% Fuzzy Logic Based System: 95.2% Decision Trees:97.93% Logistic Regression: 94.7% Precision: SVM:85.45%, ANN:99.68% BayesianNetwork:97.04% , KNN:96.84% Fuzzy Logic Based System: 86.84% Decision Trees:98.52% Logistic Regression:77.8 | Neural Network, Genetic Algorithm, Support Vector Machine, Bayesian Network, K- Nearest Neighbour, Hidden Markov Model, Fuzzy Logic Based System, Decision Trees. | --- |
Shpyrko, V. et al. [61] | ---- | Accuracy: Logistic Regress =94% Neural Network on the Sub-Sampled= 93.1% Over-Sampled= 99.9% | Logistic Regression Neural Network, Trained on Sub-Sampling, Neural Network, Trained on Over-Sampling | Cheap |
Nath Dornadula,V.et al [62] | --- | Accuracy(Before Applying SMOTE): Isolation Fores:0.9011 SVM:0.9987 Logistic Regression: 0.9990 Decision Tree:0.9994 Random Forest:0.9994 Precision(Before Applying SMOTE): Isolation Fores:0.0147 SVM:0.7681 Logistic Regression: 0.875 Decision Tree:0.8854 Random Forest:0.9310 Accuracy(After Applying SMOTE): Isolation Fores:0.5883 Logistic Regression: 0.9718 Decision Tree:0.9708 Random Forest:0.9998 Precision(Before Applying SMOTE): Isolation Fores:0.9447 Logistic Regression: 0.9831 Decision Tree:0.9814 Random Forest:0.999 | Isolation Forest, SVM, Logistic Regression, Decision Tree, Random Forest
| Expensive |
Thennakoon,A. et al. [63] | --- | Accuracy Rates : LR=74%, NB=83%, LR=72% SVM.=91% | SMOTE, RUS, CNN,SVM, Naive Bayes,KNN, Logistic Regression | Cheap |
V. Conclusion
In this paper, first, the literature of fraud, types of e-card fraud, characteristics of reliable fraud detection system, challenges of detecting e-card fraud and conventional machine learning algorithms used in detecting bank transaction frauds were presented.
By studying fraud detection researches in articles related to fraud detection, it was observed that in most cases, related supervised algorithms such as Neural Logistics Naiver Bayes, Decision Tree, Vector support and Network Regression Machine were used in implementations.
Also, by examining the algorithms and proposed methods(models) and related works and comparing the criteria such as speed of detecting fraud, accuracy and cost (of hardware and network resources), it was observed that each algorithm and methods (models) suggested related works have strengths and weaknesses. Clearly, these algorithms can be used alone, Ensemble or Meta-Learning techniques to build a stronger classifier. These techniques have been relatively successful in detecting fraud and reducing costs. However, there are still challenges such as feature engineering, parameter selection and hyper parameter, lack of sufficient data to research and change the behavior of cheaters in this field.
Obviously, combination of fraud detection techniques will be most commonly used because they combine the strengths of several detection methods and in most cases cover their weaknesses.
The main problem in the fraud detection system is the real-time identification of transactions, in which in-memory processing tools speed up performing the model. In future work, the use of real-time tools with rapid detection of patterns will reduce the financial effects and losses of fraudulent transactions.
[8] R B,A, K.K R, S.,2021. Credit Card Fraud Detection Using Artificial Neural Network, Elsevier B.V. on behalf of KeAi Communications Co, pp.35-41.
[9] Adewumi,A.O, Akinyelu, A.A.,2017. A Survey Of Machine-Learning And Nature-Inspired Base Credit Card Fraud Detection Techniques , International Journal of System Assurance Engineering and Management, pp.937-953.
[10] Roy,A., Sun,J., Mahoney,R., Alonzi, L. Adams, S., Beling, P.,2018, Deep Learning Detecting Fraud in Credit Card Transactions, Systems And Information Engineering Design Symposium (pp.129-134).IEEE.
[11] Kim,E, Lee, J., Shin, H.,Yang, H.,Cho, S., Nam, S. Song, Y.,Yoon, J.,Kim, J.,2019. Champion-Challenger Analysis For Credit Card Fraud Detection: Hybrid Ensemble And Deep Learning, Elsevier, Expert Systems With Applications, pp.214-224.
[12] Carneiro,N., Figueira, G., Costa, M., 2017. A Data Mining Based System For Credit-Card Fraud Detection In E-Tail, Elsevier - Preprinted Submitted To Decision Support Systems, pp.91-101.
[13] Zhou,H., Sun,G., Fu,S., Jiang,W., Xue,J.,2019. A Scalable Approach for Fraud Detection in Online E-Commerce Transactions with Big Data Analytics, CMC (Computers, Materials and Continua), pp.179-19.
[14] Schlegl,Th.,Seeb¨ock,Ph,Waldstein,S.M, Schmidt-Erfurth, U., Langs, G.,2017, Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery", International Conference on Information Processing in Medical Imaging (pp.146-157).
[15]
Chalapathy,R.Chawla,S.,2020, Deep Learning For Anomaly Detection: A Survey, Machine Learning (stat.ML), pp.1-50.[16] Jain,Y., Tiwari, N., Dubey, Sh., Jain, S.,2019. A comparative analysis of Various Credit Card Fraud Detection Techniques, International Journal of Recent Technology and Engineering ,pp.402-407.
[17] Meenakshi,D., Janani,B., Gayathri,S., Mrs. Indira, N., 2019.Credit Card Fraud Detection Using Random Forest, International Research Journal of Engineering and Technology(IRJET),pp.6662-6666.
[18] Shi,B., Iyengar, S.S.,2020. Mathematical Theories of Machine Learning – Theory and Applications, Springer Nature Switzerland.
[19] Moschini,G.,Houssou,R.,Bovay,J.2020.,Anomaly and Fraud Detection in Credit Card Transactions Using the ARIMA Model,engproc,pp.1-11.
[20] F.Carcilli,F., Borgne, Y.A., Caelen, O., Bontemoi, G.,2018,Streaming Active Learning Strategies For Real-Life Credit Card Fraud Detection: Assessment And Visualization, Springer International Conference On Data Science And Advanced Analytics(pp.285–300).
[21] Bansal,A.,Jain,R,.Modi,K.,2019. Big DataProcessing Using Spark in Cloud, Springer Nature Singapore Pte Ltd.
[22] Rtayli,N. Enneya,N.,2019. Selection Features and Support Vector Machine for Credit Card Risk Identification,13th International Conference Interdisciplinarity in Engineering, pp.941–948.
[23] Zhou,X., Cheng, S., Zhu, M., Gue, Ch. Zhou, S. Xu, P., Xue, Zh., Zhang, W.,2018, A state of the art survey of data min in g-basedfraud detectio n and credit scoring, International Conference on Material Engineering and Advanced Manufacturing Technology (pp.1-15).
[24] Zhou,H., Sun, G., Fu, Sh., Jiang, W., Xue, J.,2019. A Scalable Approach for Fraud Detection in Online E-Commerce Transactions with Big Data Analytics ,Computers,Materials & Continua, pp.179-192.
[26] Peharz,R.,Vergari,A.,Stelzner,K.,Molina,A.,2018. Probabilistic Deep Learning Using Random Sum-Product Networks,Cornell University, Computer Science, Machine Learning, pp.1-10.
[27] Jain,Y.,Tiwari,N., Dubey,Sh., Jain,S.,2019. A Comparative Analysis of Various Credit Card Fraud Detection Techniques, International Journal of Recent Technology and Engineering(IJRTE),pp.402-407.
[28]
Niu,X., Wang, L., Yang, X.,2019. A Comparison Study of Credit Card Fraud Detection: Supervised Versus Unsupervised, Cornell University Computer Science Machine Learning, pp.1-9.[29] Mohammadi,M., Al-Fuqaha, A., Sorour, S., Guizani, M.,2018,Deep Learning for IoT Big Data and Streaming Analytics: A Survey, IEEE Communications Surveys&Tutorials(pp.2923–2960).IEEE
[30] Carcilli,F., Borgne, Y.A., Caelen, O., Bontemoi, G., 2017,An Assessment Of Streaming Active Learning Strategies For Real-Life Credit Card Fraud Detection, IEEE International Conference on Data Science and Advanced Analytics (pp.631-639).
[31] Carcillo,F., Borgne, Y.L., Caelen, O., Kessaci, Y., Oble, F., Bontempi, G.,2021. Combining Unsupervised and Supervised Learning in Credit Card Fraud Detection", Elsevier Information Sciences,(pp.317-331).
[32] Tuor,A., Kaplan, S., Hutchinson, B.,2017. Deep Learning Unsupervised Insider Threat Detection in Structured Cybersecurity Data Streams, Associtation for the Advancement Of Artificial Intelligence, pp.224–231.
[33] Xuan,Sh., Liu,G., Li, Zh., Zheng, L., Wang, Sh., Jiang, Ch.,2018, Random Forest for Credit Card Fraud Detection, IEEE 15th International Conference on Networking(pp. 1-6).
[34] Eldin M. Abd Elhamid,H., Khalifa, W., Roushdy, M., M. Salem, A.B.,2019. Machine Learning Techniques for Credit Card Fraud Detection, Future Computing and Informatics Journal, pp.98-112.
[35] Pozzolo,A.,Boracchi,G.,Caelen,O.,Alippi, C.,Bontempi,G.,2018,Credit Card Fraud Detection:A Realistic Modeling And A Novel Learning Stratrgy, IEEE Transaction on Neural Networks And Learning Systems,(pp.3784 – 3797).
[36] Hajewski,J., Oliveira,S.,2019. Recent Advances In Big Data and Deep Learning, Springer-Proceedings Of The INNS Big Data And Deep Learning Conference.
[37] Ergen,T., Mirza, A.H., Kozat, S.S.,2017, Unsupervised and Semi-Supervised Anomaly Detection With LSTM Neural Networks, IEEE Transactions On Neural Networks And Learning Systems(pp.1-12).
[38] Zhang,X., Han, Y., Xu, W., Wanga, Q.,2021, HOBA: A novel feature engineering methodology for credit card fraud detection with a deep learning architecture, Elsevier - Information Sciences, pp.: 302-316.
[39]
Anseer,Sh, Saleem, Y., Khalid, Sh., Bashir, M.Kh., Han, J., iqbal, M.M., Han, K.,2018,Enhanced Network Anomaly Detection Based On Deep Neural Networks, IEEE Access(pp.48231–48246,.[40] Pumsirirat,A., Yan, L.,2018. Credit Card Fraud Detection Using Deep Learning Based On Auto-Encoder And Restricted Boltzmann Machine, (IJACSA)International Journal of Advanced Computer Science and Applications, pp.18-25.
[41] NOVOTNY,J., A.BILOKON,P.,GALIOTOS,A.,2020. Machine Learning and Big Data with kdb+/q, Printed in Great Britain by TJ International Ltd – Padstow – Cornwall- UK Wiley.
[42] Zamini,M. Gh.Montazer,2018,Credit Card Fraud Detection Using Autoencoder Based Clustering, IEEE 9th International Symposium On Telecommunications(pp.486-491).
[43] Gyamfi,N., Abdulai.J.,2018, Bank Fraud Detection Using Support Vector Machine, IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (pp.37-41).
[44] A.E.Hassanien, A.Darwish, H.El-Askary,2020. Machine Learning and Data Mining in Aerospace Technology, Springer Nature Switzerland AG.
[46] Zheng,P., Yuan, Sh.,Wu,X., Li, J., Lu, A.,2019, One-Class Adversarial Nets for Fraud Detection , The Thirty - Third AAAI Conference on Artificial Intelligence (pp.1286-1293).
[47] Song,H., Jiang,Zh., Men.A and Yang,B.,2017.A Hybrid Semi-Supervised Anomaly Detection Model for High-Dimensional Data, Computational Intelligence and Neuroscience,pp.1-9.
[48]
Kantardaic,M.,2019. Mining: wars, Models, Methods, and Algorithms, Wiley-IEEE Press.[49]
Bhatia,P.,2019. data mining and data warehousing principle and practical technology, Cambridge University Press.[50] Misraa,S., Thakura,S., Ghosha,M., Kumar Saha, S.,2019.An Autoencoder Based Model for Detecting Fraudulent Credit Card Transaction, Published by Elsevier Procedia Computer Science, pp.254–262.
[51] Chawla,A., Lee, B.,Fallon, Sh., Jacob, P.,2018,Host Based Intrusion Detection System With Combined CNN/RNN Model, Joint European Conference on Machine Learning and Knowledge Discovery in Databases(pp.149-158).
[52] Carcilli,F., pozzolo, A.D., borgne, Y.a., Caelen, O. Mazzer, Y., Bontemoi, G.,2018. SCARFF: a Scalable Framework For Streaming Credit Card Fraud Detection With Spark, Information Fusion 41,pp.182-194.
[53] Bharathi,T., Anuradha, S.,Phil Scholar, M., 2018. Fraud Detection Techniques on Big Data Cloud Computing, International Journal of Computer Science Engineering & Technologu(IJCSET),pp.11-16.
[54] Suman Mishra,J., Panda, S., Kumar Mishra, A. ,2013. Novel Approach for Credit Card Fraud Detection Targeting the Indian Market, IJCSI International Journal of Computer Science Issues,pp.172-179.
[55] Sohi,S.M., Ganji, F., Seifert, J.,2020. RNNIDS: Enhansing Network Instrusion Detection Systems through Deep Learning, computers & Security Elsevier, pp.1-23.
[56]
Ba,H., 2019. Improving Detection of Credit Card Fraudulent Transactions Using Generative Adversarial Networks, Cornell University, Computer Science, Machine Learning,pp.1-10.[57] TranP,H., Tran, K.P., Huong, T.T., Heuchenne, C., Huong Le, T.M.,2018, Real Time Data - Driven Approaches For Credit Card Fraud Detection,International Conference on E-Business and Applications(pp.1-4).
[58]
Sathyapriya,M., Thiagarasu, Dr.V., 2019. A Cluster Based Approach for Credit Card Fraud Detection System using Hmm with the Implementation of Big Data Technology, International Journal of Applied Engineering Research,(pp.393-396).[59] Olson,D.L., Wu, D.,2020.Predictive Data Mining Models, Springer Nature Singapore Pte Ltd.
[60] Lebichot,B., Ael Le Borgne, Y., He-Guelton, L.,.ed´eric Obl´ F., Bontempi, G.,2020. Deep-Learning Domain Adaptation Techniques for Credit Cards Fraud Detection Springer Nature Switzerland AG, pp.78–88.
[61] Shpyrko , V. , Kova, B.,2019, Fraud detection models and payment transactions analysis using machine learning, SHS Web of Conferences 65 M3E2(pp. 1-7).
[62] Nath Dornadula , V., Geetha,S.,2019, Credit Card Fraud Detection using Machine Learning Algorithms, INTERNATIONAL CONFERENCE ON RECENT TRENDS IN ADVANCED COMPUTING(pp.631–641).
[63]
Thennakoon, A., Bhagyani, Ch., Premadasa, S., Mihiranga, Sh., Kuruwitaarachchi, N.,2019,Real-time Credit Card Fraud Detection Using Machine Learning, 9th International Conference on Cloud Computing, Data Science & Engineering(pp.488-493).[64]
Zhou,H., Sun, G., Fu, S., Jiang, W., Xue ,J.,2019.A Scalable Approach for Fraud Detection in Online E-Commerce Transactions with Big Data Analytics, Computers, Materials and Continua,.p.179-192.[65]
Mazumder,S., Singh Bhadoria, R., Chandra Deka, G.,2017.Distributed Computing in Big Data Analytics Concepts, Springer International Publishing AG.[66] Al-Shabi,M.A.,2019.Credit Card Fraud Detection Using Autoencoder Model in Unbalanced Datasets, Journal of Advances in Mathematics and Computer Science,pp.1-16.
[67] https://archive.ics.uci.edu/ml/datasets/default+of+credit+card+clients
[68] Jadhav Shradha, M., Chalwa Prabhavati, V., Bakshi Asmita, S.,2020. Online Credit Card Fraud Detection System, International Research Journal of Engineering and Technology,pp.573-577.
[69] M.J.Simi,2019.Credit Card Fraud Detection : A Comparison using Random Forest, SVM and ANN, International Research Journal of Engineering and Technology,pp.225-228.
[70] S.Kumar,M., Soundarya,V., Kavitha,S., S.Keerthika, E.,Aswini,E.,2019,Credit Card Fraud Detection Using Random Forest Algorithm, 3rd International Conference on Computing and Communications Technologies,pp.149-153.
[71] ATA,O.,HAZIM,L.,2020. Comparative Analysis of Different Distributions Dataset by Using Data Mining Techniques on Credit Card Fraud Detection, Technical Gazette 27,pp. 618-626.
[72] R. Ross,Ch.,2020. Reducing Payment-Card Fraud”, Walden University.
[73] Olson,D.L.,Wu, D.,2020. Predictive Data Mining Models, Springer Nature Singapore Pte Ltd.
[74] Yazici,Y.,Approaches to Fraud Detection on Credit Card Transactions Using Artificial Intelligence Methods, pp. 235-244.
[75] Carcilli,F., Pozzolo, A.D., Borgne, Y.A., Caelen, O., Mazzer, Y., Bontemoi, G.,2018.SCARFF : a Scalable Framework For Streaming Credit Card Fraud Detection With Spark, Information Fusion 41, pp.182-194.
[76] Carcilli,F.,.Borgne, Y.A, Caelen, O.,Bontemoi, G.,2018, Streaming Active Learning Strategies For Real-Life Credit Card Fraud Detection:Assessment And Visualization , Springer International Conference On Data Science And Advanced Analytics, pp.285–300.
[77] Gupta,G.,2019. Analysis of Various Credit Card Fraud Detection Techniques , International Journal of Recent Technology and Engineering,pp.402-407.
[78] Shpyrko,V.,Kova, B,2019, Fraud detection models and payment transactions analysis using machine learning, SHS Web of Conferences 65,pp.1-7.
[79] Vidhya,K., Subhashree,.S., Swetha,S.,2021. Credit Card Fraud Detection Using Machine Learning, International Journal of Future Generation Communication and Networking, pp. 2567–2575.
[80] E. Melo-Acosta ,G.,Duitama-Mu ̃noz,F.,D.Arias-Londo ̃no,J.,2017. Fraud Detection in Big Data using Supervised and Semi-supervised Learning Techniques, pp.1-6.
[81] MisraaS.,Thakura,S.,Ghosha,M.,Kumar Saha,S., 2019. An Autoencoder Based Model for Detecting Fraudulent Credit Card Transaction,The Author(s). Published by Elsevier B,pp.254–262.
[82] Patil,G., Sakarkar,G.Dutta,P.,2021. Machine learning algorithms using Python programming ,Nova Science Publishers.
[83] Hudgeon,D.,Nichol,R.,2020.Machine Learning for Business: Using Amazon SageMaker and Jupyter` ,Manning Publications.
[84] Engels,Ch.,Kumar,K.,Philip,D.2020,Financial Literacy and Fraud Detection, The European journal of finance, pp.420-442.
[85] Sohi, S.M., Ganji , F. ,Seifert,J.,2020. RNNIDS: Enhancing Network Intrusion Detection Systems Through Deep Learning", Cryptography and Security,pp.1-23.
[86] Houssou,R.,Bovay,J.,Robert,S.,2019.Adaptive Financial Fraud Detection in Imbalanced Data with Time-Varying Poisson Processes, Journal Of Financial Risk Management,pp.1-28.
[87] Nur-E-Arefin,Md. M.Sultan,M.,2020. A Comparative Study of Machine Learning Classifiers for Credit Card Fraud Detection, International Journal of Innovative Technology and Interdisciplinary Sciences ,pp.95-406.
[88] Yang,W.,Zhang,Y.,Ye,K.,Li,L.,Zhong Xu,Ch.,2019. FFD: A Federated Learning Based Method for Credit Card Fraud Detection, Springer Nature Switzerland.
[89] Lucas ,Y., Edouard Portier ,P.,.Laporte,L.,He-Guelton,L.,Caelen,O.,Granitzer,M.,Calabretto,S.2020.Towards automated feature engineering for credit card fraud detection using multi-perspective HMMs, Future Generation Computer Systems,pp.393–402.
[90] Han,W.,Xiao,Y.,2019. Edge computing enabled non‑technical loss fraud detection for big
data security analytic in Smart Grid, Journal of Ambient Intelligence and Humanized Computing, pp.1-12.
[91] Xu,l.,2020.Neyman-Pearson Classification for Credit Card Fraud Detection, A thesis submitted in partial satisfaction of the requirements for the degree.
[92] ATA,O.,HAZIM,L.2020,Comparative Analysis of Different Distributions Dataset by Using Data Mining Techniques on Credit Card Fraud Detection, Technical Gazette 27(pp. 618-626).
[93] Yazici, Y.,2020.APPROACHES TO FRAUD DETECTION ON CREDIT CARD TRANSACTIONS USING ARTIFICIAL INTELLIGENCE METHODS, Computer Science & Information Technology,pp. 235-244.
[94] Mehana,A.,PrievaNuci,K.,2020,FRAUD DETECTION USING DATA-DRIVEN APPROACH, (pp.1-7).
[95] Mazumder,S.,Singh Bhadoria,R., Chandra Deka, G., 2017.Distributed Computing in Big Data Analytics Concepts, © Springer International Publishing AG,DOI: 10.1007/978-3-319-59834-5.
[96] SADGALI, I., SAEL, N., BENABBOU, F.,2019, Performance of machine learning techniques in the detection of financial frauds, Second International Conference on Intelligent Computing in Data Sciences(pp:45-54).
[97] Jain,Y.,Tiwari,N.,Dubey,S.,JainS.,2019.A Comparative Analysis of Various Credit Card Fraud Detection Techniques, International Journal of Recent Technology and Engineering,pp.402-407.
[98] Amanze, B.C., Asogwa, D.C., Chukwuneke, C.I.,2018. Credit Card Fraud Detection System Using Intelligent Agents and Enhanced Security Features, International Journal of Trend in Research and Development,pp.524-530.
[99] Olexander, Sh.,2021, DETECTING CREDIT CARD FRAUD USING MACHINE LEARNING ALGORITHMS, DOI 0.51582/interconf.19-20.08.2021.037 ,pp.393-403.
[100] Gamini, P., Tejasri Yerramsetti, S., Darapu, G.D.,2021.A Review on the Performance Analysis of Supervised and Unsupervised
algorithms in Credit Card Fraud Detection, International Journal of Research in Engineering, Science and Management,pp.23-26.
[101] Tiwari, P.,Mehta, S.,Sakhuja, N.,Kumar, J., Singh, A.K.,2021,CREDIT CARD FRAUD DETECTION USING MACHINE LEARNING:A STUDY”, arXiv:2108.10005v1 .
[102] Bodepudi, H.,2021. Credit Card Fraud Detection Using Unsupervised Machine Learning Algorithms , International Journal of Computer Trends and Technology,pp.1-3.
[103] Goyal, R., Manjhvar A.K.,2020. Review on Credit Card Fraud Detection using Data Mining Classification Techniques & Machine
Learning Algorithms,international Journal of Research and Analytical Reviews ,pp.972-976.
[104] Seilsepour,A., Ravanmehr,R., Sima,H,R. 2019. 2016 Olympic Games on Twitter: Sentiment Analysis of Sports Fans Tweets using Big Data Framework, Journal of Advances in Computer Engineering and Technology Volume 5, Issue 3 ,pp.129-204 .
[105] R.J, 2019. A Hybrid Data Clustering Algorithm Using Modified Krill Herd Algorithm and K-MEANS, Journal of Advances in Computer Engineering and Technology Volume 5, Issue 2 ,pp. 93-106.
[106] Khakata.E N, Omwenga.V O, Msanjila. S S, 2020. Prediction of Student Learning Styles using Data Mining Techniques, Journal of Advances in Computer Engineering and Technology, Volume 6, Issue 2, pp.107-118.
[107] Keshtkar Langaroudi.m, Yamaghani.M R, 2019. Sports Result Prediction Based on Machine Learning and Computational Intelligence Approaches: A Survey, Journal of Advances in Computer Engineering and Technology, Volume 5, Issue 1 , pp. 27-36.
Hamid Banirostam is a Ph.D. student in the Department of Computer Engineering, Central Tehran Branch, Islamic Azad University, Tehran, Iran. His research interests include Machine learning, Deep learning, Big Data, Data Analytics, Python Programming.
Touraj Banirostam is an Assistant Professor in the Department of Computer Engineering, Central Tehran Branch, Islamic Azad University, Tehran, Iran. His research interests include Cognitive Science Engineering, Artificial Intelligence, Learning, Self-Management Systems.
Mir Mohsen Pedram received the B.Sc. degree in electrical engineering from the Isfahan University of Technology, Isfahan, Iran, 1990, and the M.Sc. and Ph.D. degrees in electrical engineering from Tarbiat Modares University, Tehran, Iran, in 1994 and 2003, respectively. He is currently an Associate Professor with the Department of Electrical and Computer Engineering, Kharazmi University. His main areas of research are intelligent systems, machine learning, data mining, and cognitive science.
Amir Masoud Rahmani is currently working as a Professor for Islamic Azad University, science
and research branch, Tehran. He is the author/co-author of more than 220 publications in technical journals and conferences. His research interests are in the areas of distributed systems, wireless sensor networks, Internet of Things and evolutionary computing. Address: Amir Masoud Rahmani, Computer Engineering dept, Islamic Azad University,Science and Research branch, Hesarak, Ashrafi Esfahani,Poonak Square, Tehran, IRAN.
