Collaborative Data Cleaning Framework - a Pilot Case Study for Machine Learning Development

Aug 3, 2026

Upload by Ezra

This study experiments with collaborative data cleaning as a pivotal step in data preparation for both analysis and machine learning development. A provenance Data Cleaning Model (DCM) is employed to support multi-user scenarios, tracking dataset changes while running experiments that simulate multiple data curators collaborating on the same data. The work evaluates how distinct cleaning scenarios influence dataset completeness and correctness metrics, and how those quality changes propagate into downstream machine learning modeling performance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

AccessFree to download

Abstract

Introduction and Overview

Background and Motivation

Machine Learning Pipeline and Data Cleaning

IJDC | Conference Paper Collaborative Data Cleaning Framework: a Pilot Case Study for Machine Learning Development Nikolaus Nova Parulian University of Illinois at Urbana Champaign Bertram Ludäscher University of Illinois at Urbana Champaign Abstract This study experiments with collaborative data cleaning, a pivotal phase in data preparation for both analysis and machine learning. We used a provenance Data Cleaning Model (DCM) for multi-user scenarios to track changes on a dataset and conduct comprehensive experiments that simulate multiple data curators working collaboratively on a dataset. Furthermore, we analyzed how different data-cleaning scenarios to improve quality metrics of completeness and correctness of a dataset can affect the downstream machine learning modeling performance. Submitted 10 February 2024 ~ Accepted 22 February 2024 Correspondence should be addressed to Nikolas Parulian, [Email: nnp2@illinois.edu](Email: nnp2@illinois.edu) This paper was presented at the International Digital Curation Conference IDCC25, 19-21 February 2024 The International Journal of Digital Curation is an international journal committed to scholarly excellence and dedicated to the advancement of digital curation across a wide range of sectors. The IJDC is published by the University of Edinburgh on behalf of the Digital Curation Centre. ISSN: 1746-8256. URL: [http://www.ijdc.net/](http://www.ijdc.net/) Copyright rests with the authors. This work is released under a Creative Commons Attribution License, version 4.0. For details please see [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/) International Journal of Digital Curation 2024, Vol. 18, Iss. 1, pp. 12 1 [http://dx.doi.org/10.2218/ijdc.v18i1.924](http://dx.doi.org/10.2218/ijdc.v18i1.924)[ ](http://dx.doi.org/10.2218/ijdc.v18i1.924)DOI: 10.2218/ijdc.v 18i1.924 Introduction and Overview Data cleaning is a critical step the data preparation and aims to ensure the accuracy, consistency, and integrity of data used for analysis or machine learning. However, data cleaning is time-consuming and labor-intensive, often requiring domain expertise and manual effort. With the increasing availability of complex datasets, collaborative data cleaning has emerged as a promising approach to leverage multiple data curators’collective knowledge and skills to clean data more efficiently and effectively. Building upon our previous work on providing a provenance model for Collaborative Data Cleaning (CDCM) (Parulian et al., 2021), (Parulian and Ludäscher, 2022), in this work we describe the use case and an example of implementation of the model for a machine learning development. We aim to address the challenges and opportunities associated with collaborative data cleaning and investigate how different data-cleaning scenarios can impact the cleaned datasets and outcomes for the downstream tasks—increasing the transparency by recording provenance on each cleaning step and analyzing the effect on the data quality (DQ) metrics. To this end, we propose an experiment that involves applying different variations of datacleaning steps to simulate multiple curators working together on cleaning a dataset. Our provenance model can also support reporting queries for multi-curator scenarios such as: Who changed the values in the dataset? ; Who contributed to the data-cleaning workflow development? The findings from our experiment use case contribute to our understanding of practical use cases for collaborative data cleaning. The outcomes ofour analysis shed light on how a collaborative data cleaning process can impact the quality and reliability of the cleaned data and how it can influence downstream data analysis tasks or modeling outcomes. This study aims to advance the use of provenance for trust and transparency in collaborative data cleaning and provide valuable insights for data curators working on machine learning development. Bac

FAQs

What is the main goal of the collaborative data cleaning framework in this paper?

The paper aims to study collaborative data cleaning as a key phase in preparing data for analysis and machine learning, using a provenance model to improve transparency and understand impacts on results.

How does the provenance Data Cleaning Model (DCM) support multi-user collaboration?

The DCM tracks changes at the dataset level across multiple curators and enables experiments that simulate collaborative cleaning workflows.

Which dataset quality aspects are evaluated, and why do they matter for machine learning?

The study analyzes completeness and correctness metrics, and examines how improvements or differences in these metrics affect downstream machine learning model performance.

Aug 3, 2026

Upload by Ezra