Introduced the Anti-GraphRAG paradigm — surfacing what's absent in a knowledge graph, not just what's present.
Working on the intersection of graph machine learning, contrastive learning, and applied decision intelligence.
Where GraphRAG asks "what's connected?" — Anti-GraphRAG asks "what's missing, and does it matter?"
I introduced Anti-GraphRAG: an inversion of the standard GraphRAG retrieval paradigm that surfaces absent edges in a knowledge graph rather than present ones. The core insight is that meaningful absence — a missing drug-target interaction, an undocumented threat actor technique, an unlinked regulatory obligation — often carries more signal than what the graph already contains.
How it works:
- Construct a domain knowledge graph and freeze it at time T
- Train ComplEx embeddings on the frozen graph (link existence)
- Apply a Contrastive Void Network (CVN) — trained on temporal ground truth — to rank absent edges by significance, not just likelihood
- Validate: check which predicted voids were later confirmed by real-world data
Results on open datasets:
| Domain | Dataset | Ground Truth | P@100 |
|---|---|---|---|
| Biomedical | Hetionet + STRING v12 | Protein interactions added post-cutoff | 0.94 |
| Threat Intelligence | MITRE ATT&CK Enterprise | ATT&CK edges added post-cutoff | 0.86 |
The temporal validation method — freeze, predict, wait, verify — is what separates significance from noise. Any system can score candidate edges; this one is calibrated against what actually mattered later.
→ void-graph on GitHub · Open source · Validated on real-world knowledge graphs
Trained and evaluated KGE models (ComplEx, RotatE) for link prediction on biomedical and cybersecurity graphs. Extended standard KGE pipelines with contrastive significance heads and temporal holdout validation.
Developed the CVN (Contrastive Void Network) — an InfoNCE-trained head on frozen KGE embeddings that learns to distinguish structurally significant voids from random non-edges. Applied to Hetionet (47K nodes, 2.25M edges) and MITRE ATT&CK (1,757 nodes, 20K edges).
- Biomedical: Drug-target, gene-disease, compound-pathway void detection
- Cybersecurity: Threat actor → technique gap prediction; mitigation coverage analysis
- Open KG tooling: Knowledge graph construction and evaluation utilities
- ai-data-analyst — Natural language interface for structured data analysis
- EvoPrompt — Evolutionary prompt optimization
- FlyPyAgents — Lightweight Python agent framework
- PodcastGen — LLM-powered podcast generation pipeline
Interested in knowledge graphs, contrastive learning, or structured absence detection?