This blog post discusses the process of performing knowledge graph completion using PyKEEN and Neo4j. It explains the difference between knowledge graph embedding models and node embedding models, highlighting the importance of relationship types in multi-class link prediction. The post provides step-by-step instructions on how to transform a Neo4j graph into a PyKEEN graph and perform a train-test data split. It also demonstrates how to train a knowledge graph embedding model and predict new relationships. The results can be stored back in Neo4j for further evaluation. The post concludes by emphasizing the usefulness of knowledge graph embedding models in multi-class link prediction tasks, particularly for drug repurposing.
| Signal | Change | 10y horizon | Driving force |
|---|---|---|---|
| Knowledge graph completion with PyKEEN and Neo4j | Integration of PyKEEN with Neo4j for multi-class link prediction | Improved accuracy and efficiency in knowledge graph completion | Improve prediction capabilities and enhance data analysis |
| Use of knowledge graph embedding models for multi-class link prediction | Shift from plain node embedding models to knowledge graph embedding models | More accurate and comprehensive predictions for multi-class link prediction | Need for more accurate and nuanced predictions in link prediction tasks |
| Transformation of Neo4j to PyKEEN graph for data analysis | Neo4j graph transformed to PyKEEN graph for analysis | Simplified process for transforming and analyzing data | Streamlined data analysis and integration |
| Train a knowledge graph embedding model | Training of the RotatE embedding model for knowledge graph completion | More accurate and effective knowledge graph embedding models | Desire for improved performance and accuracy in knowledge graph embedding |
| Multi-class link prediction | Utilization of methods for predicting multiple types of links in a network | Enhanced ability to predict multiple types of links in a network | Need for predicting multiple types of links in complex networks |
| Storing predictions back to Neo4j | Storing top predictions back to Neo4j for evaluation | Improved storage and retrieval of prediction results | Facilitating evaluation and analysis of prediction results |
| Use of existing connections in the graph to explain predictions | Utilizing existing connections in the graph to assess prediction accuracy | Enhanced ability to explain and validate prediction results | Leveraging known relationships to validate predictions |
| Transformation from Neo4j to PyKEEN graph is generic and applicable to any dataset | Generic transformation method applicable to any dataset in Neo4j | Increased flexibility and applicability of the transformation process | Widening the scope of datasets that can be analyzed with PyKEEN |
| Knowledge graph embedding models are useful for multi-class link prediction tasks | Adoption of knowledge graph embedding models for multi-class link prediction tasks | Widespread use of knowledge graph embedding models for link prediction | Need for more advanced and nuanced link prediction techniques |