Towards a knowledge graph to describe and process data defects

• Joao Marcelo Borovina Josko
• Lisa Ehrlinger
• Wolfram Wöß

• Fritz Laux
• Dr. Lisa Ehrlinger

The reliability and trustworthiness of machine learning models depends directly on the data used to train them. Knowledge about data defects that affect machine learning models is most often considered implicitly by data analysts, but usually no centralized data defect management exists. Knowledge graphs are a powerful tool to capture, structure, evolve, and share semantics about data defects. In this paper, we present an ontology to describe data defects and demonstrate its applicability to build a large public or enterprise knowledge graph.