BhinnekaAI/TSB-Fraud-Tabular

Overview

TSB-Fraud-Tabular is a synthetic tabular anomaly detection benchmark developed by BhinnekaAI for evaluating machine learning methods that identify illicit financial transactions and account takeovers in digital commerce platforms.

The benchmark focuses on financial discrepancies where fraudsters attempt to evade detection by keeping transaction amounts within plausible ranges while altering behavioral combinations, such as high transaction frequencies paired with geographic anomalies during unusual hours.

Normal observations preserve expected statistical relationships between transaction amounts, account balances, and user behavior. Anomalies are introduced by disrupting these multivariate dependencies or injecting right-skewed extreme behaviors.

The benchmark is intended for research on tabular anomaly detection, financial fraud analysis, and the evaluation of global anomaly scoring methods.

Dataset information

Property	Description
Dataset name	TSB-Fraud-Tabular
Full name	Tabular Multi-Source Fraud Benchmark
Publisher	BhinnekaAI
Dataset Creator	Christopher Alberto Hamonangan
Domain	Digital banking, fintech, and financial fraud detection
Data modality	Structured numerical tabular data
Feature count	5 numerical features
Data type	Synthetic
Task	Financial fraud anomaly detection
Expected anomaly proportion	Approximately 8.0%
Label convention	0 = normal transaction; 1 = fraudulent transaction
Primary objective	Detect multivariate deviations from normal user spending profiles

Research Problem

Digital payment systems process millions of daily transactions involving amounts, user balances, geographical locations, and temporal indicators. Fraudulent transactions often do not appear as extreme outliers in individual features, but rather as inconsistent multivariate patterns.

For example, a transaction may involve a normal amount, but originate from an abnormal geographical distance during late-night hours for an account with a low balance.

Anomaly detection methods relying strictly on individual feature thresholds often fail in these scenarios. This benchmark emphasizes the detection of subtle in-cluster relational anomalies.

The benchmark investigates:

• Detection of behavioral inconsistencies in financial logs.
• Performance under right-skewed Log-Normal economic distributions.
• Evaluation of global vs. local anomaly scoring methods under imbalanced labels.

Dataset Structure

The repository follows the two-directory structure used across the BhinnekaAI HDS benchmark suite.

random-seed

Property	Specification
Directory	1_Credit_Card_Fraud_Simulator/random-seed/
Number of files	5
Observations per file	5,000
Naming convention	dataset_5000_rows_seed{seed}.csv (seeds: 11, 22, 33, 44, 55)
Intended use	Repeated-realization experiments

3.2. seed42

Property	Specification
Directory	1_Credit_Card_Fraud_Simulator/seed42/
Random seed	42
Number of files	5
File naming convention	dataset_{size}_rows_seed42.csv

Dataset size	Intended evaluation
100	Small-sample behavior
1,000	Initial model comparison
10,000	Intermediate-scale evaluation
50,000	Large-scale evaluation
100,000	Maximum specified sample-size evaluation

Feature Specification

Feature	Description	Data Type
tx_amount	Numerical transaction amount	Continuous (Log-Normal)
account_balance	Current account balance of the user	Continuous
daily_frequency	Number of daily transactions executed	Discrete (Poisson)
geo_distance	Geographical distance from home location in km	Continuous (Exponential)
night_time_indicator	Binary indicator for late-night transactions	Binary (Binomial)

Mathematical Formulation

Normal transaction amounts follow a right-skewed Log-Normal distribution reflecting standard economic expenditure:

tx_amount ~ LogNormal(μ = 5.5, σ = 0.6)

Anomalies are generated by shifting distributions toward extreme economic values while disrupting typical balance-to-amount ratios.

Loading the Dataset

import pandas as pd

df = pd.read_csv("seed42/dataset_10000_rows_seed42.csv")
X = df.iloc[:, :-1].values
y = df['is_fraud'].values
print(df.head())