FAIR Digital Object for Accessing Label Information of ML Training Data Stored in a Metadata Schema

In this poster, we introduce how the FAIR¹ Digital Object (FAIR DO) concept can simplify the access to schema-based label information of Machine Learning (ML) training data. Training data sets from heterogeneous sources mostly have different label terms. Therefore, composing them for application in ML comes with the cost of laborious relabeling. To ease this process by automation, the FAIR DO concept can be applied. A FAIR DO is an informative representation of scientific data, e.g. an ML training data set, that makes the data interpretable and actionable for computer systems. For applicability in the context of ML, a FAIR DO requires at least a globally unique Persistent Identifier (PID), mandatory metadata, and a data type. Label information of a training data set can be described using a proper metadata schema. With the self-contained structure of a FAIR DO, the associated label information can be accessed. Here we show this structure and explain how it facilitates access to label information. The latter is stored in a document that is based on a custom metadata schema built for ML. The schema provides a structure for basic label description and assignment of label terms to the elements in the training data set. With this, we point out the advantage of describing label information with a metadata schema in conjunction with the FAIR DO structure. Moreover, specialized clients and tools are needed for fully automated acting on FAIR DOs and relabeling. Extending the metadata schema with controlled vocabularies and ontologies is required for a more consistent data description. Using FAIR DOs this way could also address other laborious steps in ML training data composition like feature- or file reformatting. This work has been supported by the research program ‘Engineering Digital Futures’ of the Helmholtz Association of German Research Centers and the Helmholtz Metadata Collaboration Platform. [1] Mark D. Wilkinson et al., Scientific Data, 3, 160018 (2016)