Populating a Graph Database to Run a Usage-Based Discovery Tool

Most dataset discovery tools for Earth Observation data rely on descriptions and other metadata of the datasets, using
keyword searches or attribute filtering to determine relevance. However, these descriptions often do not include the potential
uses of the data. Thus, a user working on floods will rarely see few if any rainfall datasets show up in such a search. The
Usage Based Discovery tool, on the other hand, offers usage instances to the user, either research articles or applications,
along with the datasets that those usage instances used. This allows a user, particularly one new to the world of Earth
Observation data, to investigate which datasets are used in similar cases.

The information that powers Usage-Based Discovery is a graph database of relationships of usage to dataset and usage to
topic, allowing the user to narrow their search for similar cases. In order to scale out to a graph database rich enough to
provide a satisfactory user experience, we combine manual and automated processes to populate the graph.

The initial content of the graph has been seeded primarily via human-aided data curation methods, using sites like Google
Scholar. To scale up this effort, we’ve employed crowdsourcing. It is easy for anyone to contribute to our graph using their
Open Researcher and Contributor Identifier for authorization. We’re now experimenting with Machine Learning and Natural
Language Processing to help automate population of the graph, starting with the classification of research articles by topic.
Finding adequate training data in the absence of a comprehensive and open research article API continues to be a significant
challenge.