The main goal is to identify whether each transaction is fraudulent. Among them, the training set sample is about 590,000 (3.5% of fraud), and the test set sample is about 500,000.

The data is mainly divided into 2 categories which are joined by TransactionID. Not all transactions have corresponding identity information.

• transaction data
• identity data.

Transaction Features:

• TransactionDT: Timedelta from a given reference datetime (not an actual timestamp), but the time difference in seconds from a certain time.
• TransactionAMT: Transaction payment amount in USD, the decimal part is worth paying attention to.
• ProductCD: Product code, the product for each transaction. It may not necessarily be an actual product but may also refer to a service.
• card1-card6: Payment card information, such as card type, card category, issue bank, country, etc.
• addr1-addr2: Address, billing region and billing country
• dist: Distances between (not limited) billing address, mailing address, zip code, IP address, phone area, etc.
• P_ and (R__) emaildomain: Purchaser and recipient email domain, some transactions do not require the recipient, and the corresponding Remaildomain is empty
• C1-C14: Counting, such as how many addresses are found to be associated with the payment card, etc.
• D1-D15: Timedelta, such as days between previous transaction, etc.
• M1-M9: Match, such as names on card and address, etc.
• Vxxx: Vesta engineered rich features, including ranking, counting, and other entity relations. Some V features are missing in different proportions.

Identity Features:

• id_01-id_11: Numerical features for identity, which is collected by Vesta and security partners such as device rating, ip_domain rating, proxy rating, etc. Also it recorded behavioral fingerprint like account login times/failed to login times, how long an account stayed on the page, etc.
• DeviceType, DeviceInfo and id_12-id_38: Categorical Features

• One of the first things we noticed when conducting EDA was the sparsity of the dataset.
• The distribution of target variable 'isFraud' has class imbalance problem where it shows that 96.5% of data contains non-fraud transaction where as only 3.5% are fraud.
• Target variable 'isFraud' is more prevalent in the mobile 'DeviceType' as well as more prevalent in the 'IP_PROXY:ANONYMOUS' based on 'id_31'.
• Another observation that was immediately apparent is the imbalanced nature of the data. This shows that 'TransactionDT' is a timedelta gap, not a timestamp.
• Dataset has a very high percentage of missing values, especially the V columns.
• Anonymized columns not only had a high amount of missing data, but their distributions also were not normally distributed.

The libraries used are:

• numpy
• pandas
• matplotlib,
• seaborn
• sklearn
• lightgbm
• xgboost
• catboost

Missing Values

• Around 414 features contain missing values.
• Top features containing missing values.
• Replace NaN with a value smaller than the minimum feature value (such as -999) which is very fast and model can still find some pattern instead of losing information by dropping them.

Label Encoding

• Convert string, category, object type variables to integers type.

Outlier Removal

• Replace values with very low occurrence frequency with -9999 etc. (different from the value used for NaN replacement)

New Features

• Generate unique client identification (UID) by combining card and addr, aggregation.
• Convert time series data TransactionDT to month / week / day / time and aggregate each unit.
• Extract only decimal part (cent) of TransactionAmt.

• LightGBM
• CatBosst
• XGBoost

• The evaluation index was AUC with imbalanced data with few isFraud = 1 data.

Ensemble

Challenges:

• Sparsity of the dataset
• A lot Missing Values
• Imbalanced isFraud Variable
• Outliers

I made a wrong selection of LGBM for my final submission. I lost a medal for this competion.