Credit Card Fraud Detection: Data Analysis and Resampling Techniques

In this project, In this project, I conducted a comprehensive data analysis of credit card transactions to address the challenge of detecting fraudulent transactions. Given the highly imbalanced nature of the dataset, I focused on employing various data resampling techniques to ensure a balanced distribution of classes, thereby improving the accuracy of models in detecting fraudulent cases.

Problem Definition

Detecting fraudulent transactions with machine learning poses challenges due to class imbalance, where the majority of transactions are legitimate, and fraudulent ones are rare. Without balancing, models tend to predict the majority class, overlooking minority instances of fraud. This project explores solutions to balance the data, making it suitable for fraud detection.

Objectives

My primary objective was to analyze and balance the dataset using resampling techniques. Specifically, I applied methods like:

• Random Undersampling: Reducing the majority class to balance with the minority class.
• Random Oversampling: Increasing the minority class to match the majority class.
• SMOTE: Generating synthetic samples for the minority class to create a balanced dataset.
• ROSE: Using smoothed bootstrapping to simulate a balanced sample.

Dataset

I used the creditcard.csv dataset, containing 284,807 transactions with 30 feature columns and a target column (Class) that indicates whether a transaction is fraudulent (1) or legitimate (0).

Exploratory data analysis (EDA)

- Loading and Profiling the Dataset:

data<-read.csv(‘creditcard.csv’)

- Check dataset dimension:

dim(data)
284807     31

- Check missing value:

colSums(is.na(df))
Time 0:V1 0: V2 0: V3 0: V4 0: V5 0: V6 0: V7 0: V8 0: V9 0: V10 0: V11 0: V12 0: V13 0: V14 0: V15 0: V16 0: V17 0: V18 0: V19 0: V20 0: V21 0: V22 0: V23 0: V24 0: V25 0: V26 0: V27 0: V28 0: Amount 0: Class 0

- Check class imbalance:

table(data$Class)
     0      1 
284315    492

- Make data subset (10%):

library(dplyr)
creditcard<-data %>% sample_frac(0.1)

- Check the dimension of the subset:

library(dplyr)
creditcard<-data %>% sample_frac(0.1)

Data resampling techniques

Four data resampling techniques (random undersampling, random oversampling, SMOTE, and ROSE) are employed in this project.

Random under sampling

Random undersampling randomly selects and removes samples from the majority class. This method discards enormous amounts of data, which is quite troublesome because it can make it more difficult to learn the decision border between the minority and majority samples, which can lead to a reduction in classification performance.

library(caret)
down_train<-downSample(x=train_data, 
y = as.factor(train_data$Class))
table(down_train$Class)
 0  1 
42 42 

Random over-sampling

Random oversampling selects a random sample from the minority class and adds multiple copies of this instance to the training data. The random oversampling may increase the likelihood of overfitting occurring since it makes exact copies of the minority class samples.

library(caret)
up_train<-upSample(x=train_data,
y = as.factor(train_data$Class))
table(up_train$Class)
    0     1 
22743 22743 

SMOTE (Synthetic Minority Oversampling Technique)

SMOTE is a type of oversampling that works differently from normal oversampling. It generates synthetic data using a k-nearest neighbor algorithm. It begins by selecting random data from the minority class, and then the k-nearest neighbors are determined for this data point. It computes the vector between the current data point and the selected�neighbor using one of those neighbors. A random number between 0 and 1 is multiplied�with the vector, and this is added to the current �data point.

library(smotefamily)
smote_train<-SMOTE(train_data[,-30], train_data$Class, K=10)
smote_train<-smote_train$data
table(smote_train$class)
    0     1 
22745 22720 

ROSE (Random Over-Sampling Examples)

ROSE aids the task of binary classification in the presence of minority classes. It produces a synthetic, possibly balanced, sample of data simulated according to a smoothed bootstrapping approach.

library(ROSE)
rose_train<-ROSE(Class~., train_data)
rose_train<-rose_train$data
table(rose_train$Class)
    0     1 
11466 11319

Summary

Through this project, I gained valuable experience in:

• Handling class imbalance using different resampling techniques in R.
• Employing EDA to understand data characteristics, identify class distributions, and check for missing values.
• Preparing a balanced dataset for further analysis and modeling, demonstrating skills critical in fraud detection analysis.
• This approach to data preparation and resampling enhances the ability to develop reliable models for imbalanced data, a crucial skill in data analysis and fraud detection.

ML algorithms design and evaluation

Four different ML algorithms (logistic regression, decision tree, k-nearest neighbor, and naive bayes) are employed in this project. Each ML algorithm is evaluated for imbalanced data and four resampled data based on sensitivity, specificity, recall, the f1-score, and the ROC curve.