The Partnership’s Matching Algorithms Task Force Final Report

Ian Bogus & George Machovec

Matching Algorithms Task Force Charge

Began work in Q2 2022

Reported to the Infrastructure Working Group of the Partnership for Shared Book Collections (now the Shared Print Partnership)

• Investigate and document current algorithms used by vendors, service providers, and open source tools
• Gather input from experts
• Examine different levels of matching needs for different use cases (e.g. shared print, ILL, library analytics)
• Report back to the community

​

Early in the process commercial services were interested in the work but were mostly unwilling to share their approach to bibliographic matching due to proprietary reasons

​

Task Force Members

Current Members

• Sara Amato - EAST
• Ian Bogus (Chair) - ReCap
• Barbara Cormack - California Digital Library (CDL)
• Andy Hart- University of Pennsylvania
• George Machovec - Colo Alliance of Research Libraries
• Steve Smith - UMass Boston
• Karla Strieb - BTAA/Ohio State University
• Raiden van Bronkhorst - CDL

Former Members

David Almovodar - Pace University

Claudia Conrad (Chair) - CDL

Judy Dobry - CDL

Dana Jemison - CDL

​

Visitors

​

A variety of other people from various organizations visited including from several commercial library vendors

Why Do We Care?

Shared Print Programs

• Ensuring that retention commitments are not over or under represented

Resource Sharing

• Is the patron getting the item they requested. Bad clustering is a disaster

Collection Analytics

• Determining overlap for cooperative collection development, deselection, space analysis, etc.

Cooperative Programs

• Digitization, building communities, etc.

Rationale for What Algorithms to Study

Algorithmic Diversity: Each chosen algorithm represents a fundamentally different approach

*Organizational Transparency: The research prioritized algorithms from organizations willing to provide detailed insights into their approaches

*Vendor algorithms were not directly included for this reason

Algorithm types compared in study

Three types of algorithms were examined:

• “Match keys” based on bibliographic data
• Machine Learning/AI based matching
• Matching on standard control numbers

MARC21 Data Sets Used for Analysis

English Monographs (2013-2017)

• Current cataloging selected from two Ivy Plus libraries (with their approval)

Recent non-Roman Monographs (2013-2017)

• Current cataloging selected from two Ivy Plus libraries (with their approval)

Older English Language Monographs (pre-1950)

• Selected from two EAST libraries (with their approval)

Algorithms Overview

Gold Rush - Bibliographic Match Key Algorithm

Shared Collection Service Bus (SCSB) - Control Number Dependent Matching

MARC-AI - Machine Learning Matching Algorithms

OCLC Primary (benchmark #1) – First instance of the 035$a

OCLC Reconciled (benchmark #2) – Leveraged the WorldCat API to find merged OCLC numbers

Methodology

Library 1

Bib Records

Library 2

Bib Records

Gold Rush

Matches

MARC AI

Matches

SCSB

Matches

OCLC Primary

Matches

OCLC Reconciled

Matches

​

​

​

​

​

None of the Five Matched

Matched All

Five

Island of Uncertainty

Summary Results

Scenario 1

Recent English-language monographs (2013-2017)

Common Record Issues

• “First Edition” vs. [blank]
• Minor variations in title
• Inconsistency in the 008 dates
• Minor variations in publisher "Oxford University Press" vs "Published for the British Academy by Oxford University Press"

Scenario 2

Recent non-Roman-language monographs (2013-2017)

Common Record Issues

• Differences in record preference handling for diacritics and markings can cause problems
• Differences of where transliteration and vernacular are within records.

Scenario 3

English-language monographs (pre-1950)

Common Record Issues

• Date problems in the 008 persist.
• Missing data (e.g. records w/o authors)
• Minor variations in publisher (e.g. Wiley vs. John Wiley and Sons.)

Where to find the report

https://sharedprint.org/2025/06/13/the-matching-algorithms-task-force-releases-its-final-report/