Additionally, knowledge graphs provide fertile ground for storing data provenance related to machine learning models, their training data, and their predictions. Such metadata is invaluable for providing interpretability, which transcends mere human explanations to offer insight into why models produced certain statistical values—and how to correctly interpret them in the context of a specific use case. By inserting this information about predictions into the knowledge graph, “you can decorate the predictions with properties that describe when you did it, what model was used, the code to the training data, the algorithm,” Martin said. This information can allow you to understand when the prediction was made and what was used to make the prediction, he noted.
Synergies
The data preparation functionality of knowledge graphs amasses enterprise knowledge, connects variegated data types, and readies data for machine learning. Its analytics capabilities include machine reasoning and a number of graph algorithms for AI. “Knowledge graphs address these two different parts of the business, but in a synergistic way where the more data you have access to, the fewer chances and risks you take on the algorithmic side,” Clark concluded.