6.3.12. IEEE-CIS Fraud Detection

6.3.12.1. Summary

Real-world e-commerce transactions from the payment service provider Vesta, released for the IEEE Computational Intelligence Society’s fraud detection competition on Kaggle [1]. Each row is an online transaction, described by its amount, the card, addresses, e-mail domains, many engineered features and, for part of the transactions, the device and network used. The task is to flag fraudulent transactions.

The data is the competition’s training set, the transaction table joined to the identity table. As in Vanderschueren et al. [2], the transaction identifier and the timestamp offset are dropped.

Classes	2
Frauds	20663
Legitimate	569877
Samples	590540
Features	431

6.3.12.2. Using the Dataset

The dataset is fetched through fetch_ieee_fraud_detection. It is downloaded from OpenML on first use, which is a large download, and cached under ~/empulse_data (override with $EMPULSE_DATA_HOME or the data_home argument), so later calls work offline.

It returns a Dataset object with the following attributes:

• data: the feature matrix

• target: the target vector

• cost_matrix: a CostMatrix with default values pre-filled

• instance_costs: a dict of per-instance cost drivers ('amount')

• feature_names: the feature names

• target_names: the target names

• DESCR: the full description of the dataset

import pandas as pd
from empulse.datasets import fetch_ieee_fraud_detection

dataset = fetch_ieee_fraud_detection(backend=pd)
X, y = dataset.data, dataset.target

The backend argument selects the dataframe library used for data and target. Pass the module itself — backend=pd for pandas or backend=pl for polars.

Most features have missing values, and fitting a model on all 431 of them and 590,540 rows is slow. The example below imputes the numeric features of a sample of the transactions and trains CSBoostClassifier on them:

from empulse.metrics import Metric, Cost
from empulse.models import CSBoostClassifier
from sklearn.impute import SimpleImputer
from sklearn.pipeline import Pipeline

sample = X.sample(n=50_000, random_state=0).index
X_sample = X.loc[sample].select_dtypes(include=['number'])
y_sample = y.loc[sample]
amount_sample = dataset.instance_costs['amount'][sample]

model = Pipeline([
    ('imputer', SimpleImputer(strategy='median', keep_empty_features=True)),
    ('model', CSBoostClassifier(loss=Metric(dataset.cost_matrix, Cost()))),
])
model.fit(X_sample, y_sample, model__amount=amount_sample)

6.3.12.3. Cost Matrix

Flagging a transaction triggers an investigation with a fixed administrative cost \(c_f\), whether or not it turns out to be fraud. A fraud that is not flagged costs its full amount \(A_i\). Vanderschueren et al. [2] use this cost matrix, from Höppner et al. [3], for this dataset.

	Actual fraud \(y_i = 1\)	Actual legitimate \(y_i = 0\)
Predicted fraud \(\hat{y}_i = 1\)	tp_cost \(= c_f\)	fp_cost \(= c_f\)
Predicted legitimate \(\hat{y}_i = 0\)	fn_cost \(= A_i\)	tn_cost \(= 0\)

The investigation cost is a symbolic parameter with the default \(c_f = 10\) of both papers, exposed under the alias investigation_cost. Override it by passing the alias when evaluating the metric:

y_score = model.predict_proba(X_sample)[:, 1]

cost = Metric(dataset.cost_matrix, Cost())
default_cost = cost(y_sample, y_score, amount=amount_sample)
expensive_investigation = cost(y_sample, y_score, investigation_cost=50, amount=amount_sample)

6.3.12.4. Data Description

Most features are masked by Vesta. The groups below follow the description the competition host published [1].

Features	Description	Type
transaction_amt	Transaction amount in US dollars; also the false-negative cost	numeric
product_cd	Product code of the transaction	categorical
card1 – card6	Payment card information, such as card type, category, issuing bank and country	categorical
addr1, addr2	Billing address region and country	categorical
dist1, dist2	Distances, for example between addresses	numeric
p_email_domain, r_email_domain	E-mail domains of the purchaser and the recipient	categorical
c1 – c14	Counts, such as how many addresses are associated with the card	numeric
d1 – d15	Time deltas, such as days since the previous transaction	numeric
m1 – m9	Matches, such as between the names on the card and the address	categorical
v1 – v339	Features engineered by Vesta, including rankings, counts and entity relations	numeric
id_01 – id_11	Identity information: network connection and digital signature	numeric
id_12 – id_38, device_type, device_info	Identity information: network connection, browser, operating system and device	categorical
fraud (target)	Whether the transaction is fraudulent (1 = fraud, 0 = legitimate)	binary

Identity information is only available for about a quarter of the transactions; for the others, those features are missing.

6.3.12.5. References

[1] (1,2)

IEEE Computational Intelligence Society and Vesta Corporation. IEEE-CIS Fraud Detection. Kaggle, 2019. https://www.kaggle.com/c/ieee-fraud-detection

[2] (1,2)

Vanderschueren, T., Verdonck, T., Baesens, B., & Verbeke, W. (2022). Predict-then-optimize or predict-and-optimize? An empirical evaluation of cost-sensitive learning strategies. Information Sciences, 594, 400–415.

[3]

Höppner, S., Baesens, B., Verbeke, W., & Verdonck, T. (2022). Instance-dependent cost-sensitive learning for detecting transfer fraud. European Journal of Operational Research, 297(1), 291–300.