FAIR DO Application Case for Composing Machine Learning Training Data - Steinbuch Centre for Computing (SCC) Data Exploitation Methods (DEM)

Aug 5, 2026

Upload by Logic

Composing machine learning training data from heterogeneous sources is labor-intensive because data must be relabelled into uniform categories. The case presents how images and their label metadata can be represented as FAIR Digital Objects (FAIR DOs) so that labels and label information are easier to locate and access. Using persistent identifiers and type-based actionability, FAIR DOs support machine-actionable decisions and automate relabelling. The approach saves substantial time for scientists while enabling automated processing workflows for SEM-derived image and label datasets.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Conclusions

FAIR Digital Objects (FAIR DO) approach

PID and type-based actionability

Example: PID records for image and label metadata

Steinbuch Centre for Computing (SCC) Data Exploitation Methods (DEM) FAIR DO Application Case for Composing Machine Learning Training Data Nicolas Blumenröhr, Thomas Jejkal, Andreas Pfeil, Rainer Stotzka Composing Machine Learning (ML) training data sets from heterogeneous sources is laborious due to their relabelling into uniform categories. | | | --- | |
Different Laboratories create SEM (1) data sets, labelled heterogeneously.

Composing training data set for ML.

Laborious manual data
relabelling.

gy
biolo

cal
biologi

Label access

Label location

Label structure

| Conclusions • Images and their labels represented as FAIR DOs can be located and accessed easier. • FAIR DOs enable machine actionable decisions on the data. • This supports automated relabelling and saves a lot of time for the scientist. To automate this task, the FAIR Digital Object (FAIR DO) concept can be used. Composing ML training data from SEM (1) data sets represented as FAIR DOs. Data relabelling using PID-and type-based computer actionability for FAIR DOs. FAIR DO (2) of SEM (1) image PID information record Profile for PID Persistent Identifier (PID) | PID record of an image FAIR DO | | | | | --- | --- | --- | --- | | DO Type | DO Location | Metadata | … | | 21.11152/c700 9b0f-3cbd- 41fa-9e42-cecf0d1f3552 | [https://b2share.euda](https://b2share.euda)[t.eu/api/files/5fc88ad](t.eu/api/files/5fc88ad)[ ](t.eu/api/files/5fc88ad)[5-2f13-483c-8b80-](5-2f13-483c-8b80-)a5862c91dbbb/Biological.tar\#L7 3f9aed 9a89cf9755a6c29b2 7b069addf.jpg | 21.11152/219ca65d-a534- 4f44-aec5-a98be36f9f56 | … | resolves to FAIR DO (2) of label metadata PID resolves to points to | PID record of a label FAIR DO | | | | --- | --- | --- | | DO Type | DO Location | … | | 21.11152/de2d965b-8941-46f1-b0f0- 94e2ca41c18c | [http://mm3.datamanager.kit](http://mm3.datamanager.kit)
[.edu:8040/api/v1/metadata](.edu:8040/api/v1/metadata/)[/](.edu:8040/api/v1/metadata/)[ ](.edu:8040/api/v1/metadata/)[mldata_biology?version=1](mldata_biology?version=1) | … | record Look up the full FAIR DO PID record defines (1) Scanning Electron Microscopy (SEM) data set, provided [by R. Aversa et. al](by R. Aversa et. al). [http://doi.org/10.23728/b2share.19cc2afd23e34b92b36a1dfd0113a89f](http://doi.org/10.23728/b2share.19cc2afd23e34b92b36a1dfd0113a89f) [nicolas.blumenroehr@kit.edu](nicolas.blumenroehr@kit.edu) (2) Introduction to PIDs and FAIR DOs: [https://kit-data-manager.github.io/fairdo-cookbook/about.html](https://kit-data-manager.github.io/fairdo-cookbook/about.html)

FAQs

Why is composing machine learning training data from heterogeneous sources laborious?

It requires manual relabelling of data into uniform categories, which takes significant effort and time across heterogeneous sources.

How do FAIR DOs help make images and labels easier to locate and access?

Representing images and their labels as FAIR Digital Objects enables easier discovery and access to both the objects and their label metadata.

What role do persistent identifiers (PIDs) play in the proposed workflow?

PIDs record the FAIR DO metadata, resolve object identities, and provide machine-actionable links between image FAIR DOs and corresponding label FAIR DOs to enable automated relabelling.

Aug 5, 2026

Upload by Logic