VANILLA: Validated Knowledge Graph Completion- A Normalization-based Framework for Integrity, Link Prediction, and Logical Accuracy

🔍 Overview

VANILLA is a comprehensive framework designed to enhance Knowledge Graph Completion (KGC) by validating inferred facts using logical constraints derived from domain knowledge and normalizing the exisiting anomalies in KGs. Unlike traditional methods that rely solely on vector embeddings, VANILLA integrates symbolic reasoning with numerical learning to ensure logical validity, integrity, and accuracy of predictions. The framework uses SHACL constraints to validate predictions and supports symbolic rule mining, constraint checking, and KGC evaluation with state-of-the-art embedding models.

📁 Repository Structure

.
├── KG_Normalization/                         # KG Normalization
│   ├── KG/                         # Benchmark knowledge graphs
│   ├── French_Royalty/
│   ├── SGKG/
│   ├── SynthLC-1000/
│   ├── SynthLC-10000/
│   ├── YAGO3-10/
│   └── DB100K/
│
|   ├── Rules/                      # Symbolic horn rules for each benchmark
│   ├── French_Royalty/
│   ├── SGKG/
│   ├── SynthLC-1000/
│   ├── SynthLC-10000/
│   ├── YAGO3-10/
│   └── DB100K/
|   ├── Constraints/                # SHACL constraints
│   ├── French_Royalty/
│   ├── SGKG/
│   ├── SynthLC-1000/
│   ├── SynthLC-10000/
│   ├── YAGO3-10/
│   └── DB100K/
│
│   ├── Predictions/                # Output predictions
│   ├── LICENSE.txt
│   ├── README.md
│   ├── input.json
│   ├── symbolic_predictions_updated.py
│   ├── transform_new.py
│   └── validation.py
│   ├── Validated_KG_Completion/
│
├── KG_Normalization/
│   ├── input_KGC.json
│   ├── KGC.py
│   ├── input_KGC_hpo.json
│   ├── KGC_hpo.py

---

📊 Benchmark Statistics

KG Size	Benchmark	#Triples	#Entities	#Relations
Large	DB100K	695,572	99,604	470
	SynthLC-10000	106,549	10,000	9
Medium	YAGO3-10	1,080,264	123,086	37
	SGKG	54,585	36,450	6
Small	French Royalty	10,526	2,601	12
	SynthLC-1000	10,668	1,000	9

KG Size	Benchmark	#Constraints	#Valid	#Invalid
Large	DB100K	6	390,351	62,024
	SynthLC-10000	25	223,523	26,477
Medium	YAGO3-10	4	393,205	58,719
	SGKG	5	156,965	12,150
Small	French Royalty	2	1,922	298
	SynthLC-1000	25	22,335	2,665

---

⚙️ Setup Instructions

1. Clone the repository

   git clone [email protected]:SDM-TIB/VANILLA.git

2. Install dependencies

   pip install -r requirements.txt

3. KG_Normalization

Navigate to KG_Normalization folder and follow the steps in README of that folder

4. Validated_KG_Completion

Navigate to Validated_KG_Completion folder and follow the steps in README of that folder

📈 Evaluation Metrics

We evaluate KG completion using embedding models:

• TransE, TransH, TransD
• RotatE, ComplEx, TuckER
• CompGCN

Metrics reported:

• Hits@1, Hits@3, Hits@5, Hits@10
• Mean Reciprocal Rank (MRR)

---

🧠 Graphical Summary

The VANILLA framework integrates symbolic rules, domain constraints, and neural embeddings for high-quality knowledge graph completion. It identifies valid and invalid triples using evolving logical constraints and employs numerical models to infer missing links, ensuring semantic consistency and logical soundness in the normalized KG.

---

📄 License

This project is licensed under the terms of the LICENSE.txt.

Authors

VANILLA has been developed by members of the Scientific Data Management Group at TIB, as an ongoing research effort. The development is co-ordinated and supervised by Maria-Esther Vidal. We strongly encourage you to report any issues you have with VANILLA. Please, use the GitHub issue tracker to do so. VANILLA has been implemented in joint work by Disha Purohit, and Yashrajsinh Chudasama.