Dynamic knowledge graph embedding

Abstract

The goal of this study was to explore existing knowledge graph embedding techniques to discover one suitable for the implementation of online learning where embeddings in the model can be updated with any changes in the knowledge graph without the need to retrain the model again from scratch. Embedding a knowledge graph allows useful tasks such as prediction of links and classification of entities to be performed and this can lead to discovery of new facts within the knowledge graph. Many knowledge graphs are also updated frequently with new up to date information. However, current embedding techniques do not support online learning and having to retrain the model with every minute change in the knowledge graph is not time efficient and limits how useful each model is. Finding a way to implement online learning to a model should allow for significant time savings for updating of embeddings each time the knowledge graph is updated. We started by exploring and understanding different embedding techniques to determine which would lend themselves best to certain implementations of online learning. We then did some experimentation with selected models to test the feasibility of online learning. We gathered the design for each model from the paper released proposing the method and the implementation from their respective GitHub. We discovered a potential candidate in R-GCN that could lend itself to implementation of online learning given the correct dataset and some testing and tweaking in the future for better overall performance. More experimentation and development is required to test our current initial findings.