Updating Knowledge Graph Embeddings by Intermediate Estimations on Numerical Attributes

Graph embedding methods are used to create representations of knowledge graph entities in a high-dimensional vector space. These embeddings can be used in downstream tasks such as classification or link prediction. However, typically graph embedding methods require training on the entire knowledge graph, making them less efficient when a knowledge graph is expected to be dynamic, such as with IoT measurements graphs, which are updated throughout the day with new numerical measurements. This paper introduces a method for efficiently creating embeddings for new knowledge graph entities, without retraining the embedding model. The proposed method estimates an embedding for new entities, by averaging the embedding of the k nearest neighbors, where nearness is based on numerical attributes associated with the entities. We investigate the performance of this method, both on a synthetic knowledge graph, and five real-world knowledge graphs. In these experiments, we employ RDF2vec as the embedding method, and classification as the downstream task. We compare three distance measures for determining nearest entities. We observe a trade-off between the accuracy and efficiency of estimating the embeddings and retraining the embedding model. Resulting in a significant decrease of training time, compared to retraining the full model, with only a relatively small reduction in precision. Results show that in cases where the attributes are representative enough our method is an effective and efficient method to incrementally adjust graph embeddings.