- numpy
- pandas
- seaborn
- matplotlib
- sklearn
- This dataset contains around 290,000 credit card transaction details.
- There are 32 features in total for the dataset including the time when the transaction happened and the amount.
- According to the organization restrictions the name of the features are kept confidential.
there are no null value in the data set to fill
There are too many features so to prevent from overfitting unnecessary features are dropped. By plotting the heatmap and FacetGrid plots , the features that are less correlated to the class are dropped.
-> The features V1, V3, V5, V6, V7, V9, V10, V12, V13, V14, V15, V16, V17, V18, V23, V24 and Time are not correlated to the Class.
As they are around 290,000 examples, the data set was splitted into 90% train 10% test sets.
This data is trained on several models :
- XGBoost
- SVM
- RandomForest
- k nearest neighbour
- Logistic regression
=> As this data set contains very few exaples of class 1 (around 500) it is not a good idea to use accuracy as evaluation metric.
=> It is very important to detect the fraud transaction so the there are 2 evaluation matrics used
- model with recall for class 1 should be minimum 80%
- model with highest accuracy
| MODEL | Class 1 RECALL | ACCURACY |
|---|---|---|
| XGBoost | 0.69 | 99.95% |
| SVM | 0.64 | 99.94% |
| RandomForest | 0.82 | 98.45% |
| k nearest neighbor | 0.69 | 99.95% |
| Logistic regression | 0.85 | 97.31% |