A binary classification model that can classify whether a credit card data is fraudulent or legit.
Data has already available here. The dataset contains transactions made by credit cards in September 2013 by European cardholders. Dataset presents transactions that occurred in two days, which contains 492 frauds transactions out of 284,807 transactions. The dataset is highly imbalanced.
Initially datasets contains 1081 duplicate rows. After removing those it reduces to 283,726 observations, where positive (fraudulent) class contains only 0.001667% of data and negative class contains 0.998333% of data. To reduce imbalanced property, under and over sampling has done where over sampling performs well. On the other hand, under sampling perfomrs poorly.
Datasets is trained using Decision Tree, Random Forest and Gradient Boosting classification models. Showing each models performance, their training time and lackings.
In the table I showed the Precision, Recall, F1 score and accuracy for three models.
| Model | Precision | Recall | F1-score | Accuracy | ||||
|---|---|---|---|---|---|---|---|---|
| fraudulent | Non-fraudulent | fraudulent | Non-fraudulent | fraudulent | Non-fraudulent | Training | Testing | |
| Decision Tree | 0.53 | 1.00 | 0.53 | 1.00 | 0.53 | 1.00 | 1.00 | 0.9993 |
| Random Forest | 0.94 | 1.00 | 0.53 | 1.00 | 0.68 | 1.00 | 1.00 | 0.9997 |
| Gradient Boosting | 0.03 | 1.00 | 0.80 | 0.99 | 0.07 | 0.99 | 0.9849 | 0.985 |
Most important feature for training Decision Tree, Random Forest and Gredient Boosting is V14.
Random Forest is fitting 2 folds for each of 12 candidates, totalling 24 fits. On the other hand, Gredient Boosting is fitting 2 folds for each of 9 candidates, totalling 18 fits.
Random Forest performed best for max depth = 30 and n estimators = 75 where n estimators was 25, 50, 75 and max depth was 10, 20, 30, 40. On the other hand, Gredient Boosting performed best for max depth = 4 and n estimators = 30 where n estimators was 20, 25, 30 and max depth was 2, 3, 4.
Mean Test Score of Gradient Boosting is lower than Random Forest. But Mean Fit Time of Gradient Boosting is higher than Random Forest.
From the image we can see there is a Positive Correlation between Training Time and Estimators for Random Forest Classifier.
