An evaluation of the impact of training data locality on the effectiveness of knowledge graph embeddings models

As the volume and variety of data that many modern organisations deal with continue to grow, graphs are becoming increasingly important and relevant as a means of organising this data. This work looks at a possible way to improve the training of some state-of-the-art machine learning models in the area of knowledge graph embeddings. Where the interest of the user is on the ability to predict the existence of a particular link type as opposed to predicting links generally, subsets or sub-graphs could possibly be used to train the model more effectively than the entire graph. We evaluate the performance of two state-of-the-art knowledge graph embedding models on the task of predicting a specific link type. The models are first trained with all of the available training data and subsequently with subsets or sub-graphs based on the locality of the link type we wish to predict. We find that there is evidence that using less training data can in some cases actually improve the performance of the model. Finally, we look at some graph features and examine if there is any correlation between these and the accuracy/performance of the machine learning models. While no strong correlation is found, the results point to further work being required to understand this phenomenon.