A closer look / Case study
GNN + GAT: Knowledge Representation Research
Accuracy
F1-Score
GC4T 2025
Cora / PubMed / DBLP
The question
Existing GNN architectures struggle with multi-relational data and explainability in healthcare knowledge graphs, a gap between graph representation research and clinical application.
What I owned
Designed MetaGAT model architecture, ran benchmarks across Cora, PubMed, DBLP datasets, wrote and submitted the paper, presented at conference.
What changed
86.4% accuracy, 85.1% F1-score. Best Paper Award at GC4T 2025. Applied framework to disease prediction and recommender systems.
The tradeoffs
Prioritised benchmarkable accuracy on standard datasets over clinical validation (which would require IRB-level access). Explainability layer is v2 research.
“The MetaGAT architecture addressed a problem the community has been circling for a while, multi-relational graph learning with interpretability built into the design rather than added after. The benchmark results held up across all three datasets.”
The Outcome
Awarded Best Paper at GC4T 2025 by IEEE Pune Section. In the reported study, the MetaGAT framework reached 86.4% accuracy on multi-relational graph prediction; its use in healthcare recommenders and disease prediction remains a direction for further validation.