From implementation to application: FAIR digital objects for training data composition

• DOI: 10.3897/rio.9.e108706
• Publication Date: 2023-08-22T06:46:06Z
• Authors: Nicolas Blumenröhr; Rossella Aversa
• Keywords: ddc:004; Machine Learning; Science; Linked Data; DATA processing & computer science; Q; Operations; Metadata Schemas; Vocabulari; info:eu-repo/classification/ddc/004; FAIR Digital Objects; 004; Vocabularies; Python (programming language); Data Reuse; Data Sharing; Proof of concept

Composing training data for Machine Learning applications can be laborious and time-consuming when done manually. The use of FAIR Digital Objects, in which the data is machine-interpretable and -actionable, makes it possible to automate and simplify this task. As an application case, we represented labeled Scanning Electron Microscopy images from different sources as FAIR Digital Objects to compose a training data set. In addition to some existing services included in our implementation (the Typed-PID Maker, the Handle Registry, and the ePIC Data Type Registry), we developed a Python client to automate the relabeling task. Our work provides a Proof-of-Concept validation for the usefulness of FAIR Digital Objects on a specific task, facilitating further developments and future extensions to other machine learning applications.