IEEE-CIS_Fraud_Detection

Predicting the probability that an online transaction is fraudulent, as denoted by the binary target isFraud.
The dataset is from Kaggle, containing 59,0540 rows and 433 features.

Validation Strategy

The timespan of the total data set is 365 days, where that of the training set is 182 days and that of the test set is 183 days. Thus, the validation strategy used in this project is time-based validation, training for the first 5 months and predicting the last month.

EDA

Data Processing

Dimension Reduction

• Apply PCA to highly correlated and redundant V1-V339 features

Feature Selection

• Perform Adversarial Validation to find features that are important in differentiating cards as the training set and test set have different sets of cards

Feature Engineering

• Combining features to generate new features
  Feature A and B by themselves may not correlate with the target variable but FeatureA+B may correlate with the target variable.

• Frequency Encoding for categorical features
  Replace categorical values with corresponding frequency

• Group statistics
  For example, group by card1, get mean or std of TransactionAmt for each group. This can let the model know whether a row has abnormal TransactionAmt for their group.

Process TimeDelta Features

• Normalize D columns to prevent them from increasing by time
• Convert TransactionDT into datetime by providing a reference datetime

Mapping Emails to keep only Email Domain

Model Training

• Parameter Tuning with Hyperopt
• XGBoost (ROC_AUC: 0.9280)
• CatBoost (ROC_AUC: 0.9146)