PartKG2Vec: Embedding of Partitioned Knowledge Graphs

Abstract

AbstractLarge-scale knowledge graphs with billions of nodes and edges are increasingly common in many domains. Such graphs often exceed the capacity of the systems storing the graphs in a centralized data store, not to mention the limits of today’s graph embedding systems. Unsupervised machine learning methods can be used for graph embedding, which can then be used for various machine learning tasks. State-of the art embedding techniques are often unable to achieve scalability without losing accuracy and efficiency. To overcome this, large knowledge graphs are frequently partitioned into multiple sub-graphs and placed in nodes of a distributed computing cluster. Graph embedding algorithms convert a graph into a vector space where the structure and the inherent property of the graph is preserved. Running such algorithms against these fragmented sub-graphs poses new challenges, such as maximizing the likelihood of preserving network neighborhood of nodes. Also, the learned embeddings of the individual graph partitions need to be merged into one overall embedding to maximize the likelihood of preserving network neighborhood of nodes. This paper introduces a novel method for embedding of partitioned knowledge graphs. It partitions the knowledge graph and executes learning algorithm in parallel on the partitions and merge their outputs to produce an overall embedding. Our evaluation demonstrates that the runtime performance is improved after partitioning of knowledge graph against complete knowledge graph and the quality of the embedding is like that of an embedding produced on the complete, unpartitioned graph.KeywordsKnowledge graphsGraph partitioningFeature learningNode embeddingGraph representation learning