Credit Card Fraud Detection

METHOD 1 RNN

Illustration of the Architecture

Data Preprocessing

Numeric Feature

['locdt', 'loctm', 'conam', 'iterm', 'flam1', 'csmam']

Step 1. Convert string to integer

Step 2. Normalization

Category feature

['chid', 'cano', 'contp', 'etymd', 'mchno', 'acqic', 'mcc', 'ecfg', 'ins fg', 'bnsfg', 'stocn', 'scity', 'stscd', 'ovrlt', 'flbmk', 'hcefg', 'csmcu', 'flg_3dsmk']

Step 1. Convert string to token

Step 2. Embedding (vector size = min(16, vocabulary size))

Training

Step 1. Embedding

Step 2. Concatenate numeric and embedded category features

Step 3. RNN forward pass (LSTM)

Step 4. Fully connected layer for prediction

TODO

1. Prediction

   (1) Threshold

   (2) Voting

2. Accelerate

   (1) Adjust parameter, e.g. hidden size, train batch, vector size, etc.

METHOD 2 XGBoost

Data Preprocessing

Numeric Feature

['locdt', 'loctm', 'conam', 'iterm', 'flam1', 'csmam']

Step 1. Convert string to integer without normalization

Category feature

['chid', 'cano', 'contp', 'etymd', 'mchno', 'acqic', 'mcc', 'ecfg', 'ins fg', 'bnsfg', 'stocn', 'scity', 'stscd', 'ovrlt', 'flbmk', 'hcefg', 'csmcu', 'flg_3dsmk']

Step 1. Replace it with the probability of fraud.

Training

$ cd XGB
$ python train.py

Inference

$ python predict.py