Baris Yapici

Brokering Data Collaborations through Knowledge Graph Representation of Funding Ecosystems

Baris Yapici summarizes his Master's thesis in which he used Knowledge Graphs to identify promising data collaborations between researchers and research projects.

Background and Motivation

Publicly funded projects generate valuable datasets and other research assets, but the information needed to find them—and the people behind them—is scattered across portals and unstructured documents. In practice, collaboration among researchers and research projects still depends heavily on chance. The question driving my thesis is therefore straightforward: how can we make research data collaboration brokerage systematic?

A project-centric approach

Many discovery systems focus on publications or individual researchers. This work instead uses the research project as the anchor entity, because projects connect intent and topic (objectives/keywords), partners (organisations), funding context, outputs and assets, as well as contact points in one administrative unit.

I implemented an automated, schema-driven pipeline (see Figure 1) that turns CORDIS (Community Research and Development Information Service) funding records into a versioned Neo4j knowledge graph (KG). Each run produces a self-contained snapshot (CSV exports plus metadata and provenance) and can be executed for a full programme or a curated project list.

Figure 1: Overview of the pipeline and how it supports brokerage decisions.

Linking to persistent identifiers (PIDs)

To connect the portfolio to the wider research ecosystem, the pipeline enriches records with persistent identifiers:

  • ROR (Research Organization Registry) matching for organisations (disambiguation + canonical identifiers)
  • DOI (Digital Object Identifier) enrichment (via Crossref) to retrieve licence signals and author lists for outputs
  • ORCID (Open Researcher and Contributor ID) captures where available
  • Coordinator contacts from IHI (Innovative Health Initiative) / IMI (Innovative Medicines Initiative) factsheets (including industry vs. academic categorisation)

Across two large IHI/IMI portfolio snapshots (ongoing and closed projects), ROR matching reached 97–98% of organisations. Asset DOI coverage was ~53–56%, while ORCID coverage for people was ~19–25%—useful, but still incomplete, and therefore a measurable “metadata gap” for real brokerage use cases.
 

Surfacing hidden assets in DMP PDFs (LLM-assisted)

Many datasets and registries are described only in data management plans (DMPs). I added an optional enrichment step that identifies DMP candidates, downloads the linked PDFs, and uses a large language model (LLM) to extract only in-scope assets (datasets, registries, databases, knowledge graphs). Every extracted entry is traceable to the source PDF and marked as LLM-generated.

In the two IHI/IMI snapshots, the system processed 41 + 51 DMP PDFs and extracted 267 and 307 unique asset records (post-filter). Hidden assets appeared in roughly a quarter of projects, showing that important data is often documented but not systematically indexed.
 

What this enables: brokerage signals and briefs

With projects, partners, topics, and assets represented in one graph, the system supports two complementary brokerage signals:

  • Pathway connectivity (e.g., shared participating organisations that already provide collaboration channels)
  • Interest alignment (transparent text similarity over project descriptions/keywords; validated by correctly identifying known “sister projects”)

For each candidate match, the pipeline can generate a structured “brief” that summarises shared organisations, relevant terms, asset inventory (including hidden DMP assets), and actionable contact points.
 

Limitations and next steps

Real-world metadata remains incomplete, and scoring functions can be improved or replaced with more advanced methods of quantitative analysis. The next step is a user-facing interface that makes these briefs easy to explore in day-to-day brokerage workflows.
 

Data and code availability

Pipeline, schema, snapshots, and a Cypher query catalogue are archived on Zenodo: https://doi.org/10.5281/zenodo.18612891