Credit Card Fraud Detection is a highly unbalanced dataset consisting of 492 fraud and 284,807 normal transactions. Every record in data frame has 30 features which are:
- Time: the number of seconds elapsed between each transaction and the first transaction;
- Amount: the transaction amount;
- V1, V2, ..., V28: generated by PCA features of transaction.
Residual neural network was built as classifier. The NN consists of 6 fully connected layers and 2 shortcuts connections. Here is its structure:
___________________________________________________________________
Layer Output Shape Activation Connected to
===================================================================
input [(None, 29)]
___________________________________________________________________
dense_0 (None, 64) ReLu input
___________________________________________________________________
dense_1 (None, 64) ReLU dense_0
___________________________________________________________________
add_0 (None, 64) dense_1
dense_0
___________________________________________________________________
dense_2 (None, 64) ReLU add_0
___________________________________________________________________
dense_3 (None, 29) ReLU dense_2
___________________________________________________________________
dense_4 (None, 29) ReLU dense_3
___________________________________________________________________
add_1 (None, 29) dense_4
dense_3
input
___________________________________________________________________
dense_5 (None, 29) ReLU add_1
___________________________________________________________________
dense_6 (None, 1) Sigmoid dense_5
===================================================================
The model is saved in model.h5 file.
Test accuracy of the model is 0.93, ROC AUC is 0.97 and F1-Score is 0.93.
The project is completed using DVC. There are 3 stages in total:
- stages/split.dvc - Pre-process the dataset: removes column Time, creates RobustScaler and use it on column Amount, split dataset into train and test with test_size=0.2;
- stages/train.dvc - Trains the model and saves it to data/model.h5 for DVC and to model.h5 for Git;
- stages/evaluate.dvc - Evaluates the model via test dataset. Outputs accuracy, ROC AUC score
and F1-Score. Saves all metrics to metrics.txt, so later you can view them using
dvd metrics show.
There is a visualisation of whole pipeline of stages/evaluate.dvc below:
+-------------------------+
| data\creditcard.csv.dvc |
+-------------------------+
*
*
*
+------------------+
| stages\split.dvc |
+------------------+
*** ****
**** ***
** ****
+------------------+ **
| stages\train.dvc | ****
+------------------+ ***
*** ****
**** ****
** **
+---------------------+
| stages\evaluate.dvc |
+---------------------+
To reproduce the model and the evaluation process use: dvc repro stages/evaluate.dvc.