|
| Cited Used | Data Description | Applied
Method | Validation
Method | Advantage | Limitation |
|
| [27] | Credit and debit card transactions of the Spanish bank BBVA, from January 2011 to December 2012 | Multilayer perceptron (MLP) | True Positive Rate (TPR), Receiver Operating Characteris-tic (ROC curves) | The improvement of accuracy in detecting results by using parenclitic networks reconstruction for feature extraction | Requires the cases where the features are not correlated or not extracted by parenclitic networks |
| [28] | Transactions of National banking group of Italy, from April 2013 to August 2013 | Multi-objective genetic algorithm | TPR, ROC
curves | Provide feature selection process via the auto-tuning method | Should be applied to cases other than Banksealer |
| [29] | UCSD Data Mining Contest 2009 data | Deep neural
network (DNN) | Mean Squared Error (MSE),
Root Mean Squared Error (RMSE),
Mean Absolute Errors (MAE),
Root Mean Squared Log Error (RMSLE) | Study on the importance of features based on the deep learning method | Does not have accurate experimental explanation process and validation |
| [30] | Dataset achieved from the second robotic & artificial intelligence festival of Amirkabir University | Decision trees | F-Measure
| Ensemble
classification is performed using cost-sensitive decision trees in a decision forest framework | Having a class imbalance problem |
| [31] | German dataset (which has been used in KDD99
competition), Australian credit cards’ open dataset | Particle swarm optimization (PSO), Teaching-learning- based optimization (TLBO) | Confusion Matrix (True positive, True negative, False positive, False negative) | Experiment with various datasets
| Detection accuracy is relatively low |
| [32] | Not specific | Linear Regression,
Artificial Neural Networks (ANN),
K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Stump, M5P Tree, Decision Table | Normalized Root Mean Squared Error (NRMSE), TPR,
F-Measure | A comparative study using various algorithms | Detection accuracy should be increased |
| [33] | UCI German
credit card dataset | SVM | K-fold
Cross validation | As Gaussian kernels including RBF are with appropriate regularization, it guarantees a globally optimal predictor which minimizes both the estimation and approximation errors of a classifier | There is no comparison with other algorithms and there is no explanation to verify SVM algorithms superiority to others |
| [34] | Credit card transaction data from commercial bank in China | Convolutional Neural Networks (CNN),
K-Means | F-Measure | Designing a feature called trading entropy based on the latest consumption preferences for each customer and generating synthetic fraudulent samples from real frauds by a cost-based sampling method | Detection
accuracy should be increased |
| [35] | Banking transaction dataset in Iran | KNN | Accuracy,
Re-call,
Precision | A novel approach combining K-nearest neighbor, association rules like Apriori algorithm | The validation is not specific and it is difficult to compare the proposed results with other algorithms |
| [36] | Open dataset: ccFraud | NN, PSO,
Auto-associative neural network (AANN), Particle swarm optimization auto-associative neural network (PSOAANN) | MSE, Classification Rate (CR) | Combined parallelization of the auto-associative neural network in the hybrid architecture | Dataset is highly unbalanced and detection accuracy should be increased |
| [37] | Open dataset:
MIT Human Dynamics Lab | SVM, Fuzzy clustering | TPR, FPR, ROC
curves | Divide the fraud detection system into two principal modules | Would be better to compare it with more diverse algorithms |
| [38] | Transactions from a large national bank in Italy,
collected from December 2012 to August 2013 | Principal component analysis (PCA),
DBSCAN | ROC
curves | Operate in online processing | Accuracy by validation is not constant |
| [39] | Not specific | Self- organizing map (SOM) | TPR, FPR | Division of transactions to form an input matrix and ability to be applied to a large complex set | Only have compared to one algorithm |
|
