Urwashanza/EuroProcure-10-ML-Fraud-Risk

EuroProcure-10-ML: A Benchmark Dataset for Procurement Fraud Risk Detection in European Public Procurement

Overview

EuroProcure-10-ML is a benchmark dataset for machine learning based procurement fraud risk detection in European public procurement. It contains 204,752 post award procurement records collected from Tenders Electronic Daily (TED), the official procurement platform of the European Union, covering the period 2016 to 2025.

The dataset is designed for research in procurement fraud detection, anomaly detection, risk assessment, explainable artificial intelligence, and public sector machine learning.

Authors / Creators

Khalid, Urwa Binat (Researcher) https://zenodo.org/search?q=metadata.creators.personororg.name:%22Khalid,+Urwa+Binat%22

Dataset Domains

The dataset spans multiple procurement domains, including:

• —Healthcare
• —Digital technology
• —Construction and infrastructure
• —Automotive
• —Green energy
• —Defense and security
• —Aerospace

Dataset Statistics

Data Source and Provenance

Original Data Source

The underlying procurement notices were sourced from Tenders Electronic Daily (TED), the official procurement platform of the European Union.

TED provides public information on European public procurement notices and is operated by the Publications Office of the European Union.

Source: Tenders Electronic Daily (TED), European Union https://ted.europa.eu/

Dataset Construction

EuroProcure-10-ML was constructed from procurement records obtained from TED and subsequently processed and enriched for machine learning research.

The dataset construction includes:

1. 1.Collection of procurement records from TED
2. 2.Structuring and organization of procurement information
3. 3.Procurement domain categorization
4. 4.Extraction and organization of CPV information
5. 5.Processing of procurement and award attributes
6. 6.Construction of procurement related features
7. 7.Engineering of theory informed procurement risk signals
8. 8.Construction of a weakly supervised fraud risk framework

The resulting dataset contains structured procurement information together with engineered risk indicators.

Dataset Features

Each record contains structured procurement information including:

• —Buyer information
• —Procurement procedures
• —Contract characteristics
• —CPV classifications
• —Funding indicators
• —Award information
• —Bidder participation measures
• —Procurement descriptions
• —Technical requirements
• —Eligibility requirements
• —Procurement risk indicators

The dataset contains 63 columns.

Fraud Risk Signals

To support reproducible research, the dataset includes engineered risk signals derived from procurement audit and procurement risk literature.

The five primary risk indicators are:

These signals are aggregated into a weakly supervised procurement fraud risk framework.

The resulting fields include:

• —risk_score
• —risk_label

Important Note About Risk Labels

The risk_label should be interpreted as a weakly supervised fraud risk indicator, not as a confirmed legal finding of procurement fraud.

The dataset identifies procurement records exhibiting combinations of predefined risk signals. A high risk label therefore indicates elevated risk according to the dataset's methodology and should not be interpreted as proof that fraud occurred.

Data Quality

Data quality was prioritized throughout dataset construction.

Missing values were preserved whenever information was unavailable in the original procurement notices. No synthetic records, fabricated observations, or artificial imputations were introduced. Users should interpret missing values as reflecting source data availability rather than processing errors.

Intended Use

The dataset is intended for research and experimentation in:

• —Procurement fraud detection
• —Fraud risk assessment
• —Anomaly detection
• —Public procurement analytics
• —Explainable artificial intelligence
• —Risk assessment
• —Public sector machine learning
• —Weakly supervised learning
• —Tabular machine learning
• —Explainable risk modeling

Dataset Format

The Hugging Face version is provided in Apache Parquet format, selected to provide efficient storage and access for a large tabular dataset.

The dataset contains 204,752 records, 63 columns, and covers 2016 to 2025.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset(
    "Urwashanza/EuroProcure-10-ML-Fraud-Risk"
)

print(dataset)

Citation

DOI: 10.5281/zenodo.21355728

Raw data sourced from TED (Tenders Electronic Daily), European Union. Enriched and processed by Urwa Binat Khalid.