End-to-end knowledge graph for Explainable AI research. Property graphs (Neo4j) + semantic web (RDF/OWL) + KG embeddings (PyKEEN), plus a working GraphRAG pipeline benchmarked against vanilla RAG. Built on 3,907 arXiv papers.
| Layer | Metric | Value |
|---|---|---|
| Embeddings | TransE baseline MRR / Hits@10 | 0.220 / 0.414 |
| Embeddings | Citation prediction precision@10 | 0.90 |
| Reasoning | SHAP-authors via OWL property chain | 4,009 (vs 0 without reasoning) |
| GraphRAG vs RAG | Win rate / completeness gap | 6/10 · +1.9 of 5 |
| Data | Papers / authors / venues / topics | 3,907 / 13,933 / 733 / 30 |
All numbers reproducible from the notebooks.
4 node types · 4 relationship types · 11 Cypher queries covering aggregations, multi-hop joins, and per-group ranking.
Notebook: neo4j_cypher_queries.ipynb
Custom ontology with FOAF/Dublin Core alignment. 87K explicit triples → 174K after OWL-RL reasoning. SPARQL queries match the Cypher ones, plus a reasoning showcase: "Who works on SHAP?" returns 0 authors without reasoning, 4,009 with the authored ∘ about → worksOn property chain active.
Notebook: rdf_sparql_queries.ipynb
| Model | Loss | Hits@10 | MRR |
|---|---|---|---|
| TransE | MarginRankingLoss | 0.414 | 0.220 |
| TransE | NSSALoss | 0.360 | 0.168 |
| RotatE | MarginRankingLoss | 0.237 | 0.130 |
| RotatE | NSSALoss | 0.364 | 0.197 |
Simplest config won. Strong model × loss interaction: aggressive training (NSSALoss + 64 negatives + 200 epochs) hurts TransE (−0.052 MRR) but helps RotatE (+0.067). Architectural complexity and training recipe must match.
Citation link prediction on arXiv:2105.07190 using RotatE+NSSALoss → 9/10 predicted citations match actual → precision@10 = 0.90.
Notebooks: train_transe_baseline, train_transe_tuned, train_rotate_baseline, train_rotate_tuned, link_prediction
Natural language → Cypher → graph results → LLM answer. Three-layer safety (prompt rule → regex validator → Neo4j read-only mode); 5/5 destructive prompts refused.
10 questions, same generator LLM (Llama 3.3 70B via Groq) for both pipelines:
| GraphRAG | Vanilla RAG | |
|---|---|---|
| Win count | 6 | 4 |
| Completeness | 4.4 | 2.5 |
| Specificity | 4.4 | 2.9 |
GraphRAG wins on structural questions (counts, multi-hop, rankings); RAG wins on conceptual synthesis. The judge showed a verification bias on factual queries — penalised correct database-verified answers it couldn't validate from retrieved abstracts (details in comparison notebook).
Q: Find papers that cite Grad-CAM and are about Healthcare applications. A: Top 5 by citation count: "Explainable AI in deep learning-based medical image analysis" (1,012 cites), "XAI Opportunities and Challenges Survey" (749), … (LLM-generated Cypher + answer in notebook)
Notebooks: graphrag.ipynb · vanilla_rag.ipynb
Neo4j Aura · rdflib + owlrl · PyKEEN · sentence-transformers (multilingual-e5-base) · Llama 3.3 70B via Groq · Python
git clone https://github.com/Pra0809/xai-knowledge-graph.git
cd xai-knowledge-graph
pip install -r requirements.txt
cp .env.example ~/.env_xai_kg # fill in Neo4j Aura + Groq keysRun notebooks in order: data_ingestion → neo4j_cypher_queries → rdf_sparql_queries → train_transe_baseline → train_transe_tuned → train_rotate_baseline → train_rotate_tuned → link_prediction → graphrag → vanilla_rag
- 3,907-paper corpus and 10-question evaluation are indicative, not statistically rigorous.
- Author dedup misses name variants (e.g., "W. Samek" vs "Wojciech Samek"); RotatE embeddings surface this but don't fix it.
- LLM-as-judge has a known verification bias — raw outputs in
data/rag_comparison_results.jsonshould be read alongside the numeric scores.
Built alongside the Foundations of Knowledge Graphs module at Paderborn University, M.Sc. Computer Science. Data: arXiv (CC BY 4.0), Semantic Scholar Academic Graph.