Semantic Preserving Embeddings for Generalized Graphs

In this work we present a new approach to the treatment of property graphs using neural encoding techniques derived from machine learning. Specifically, we will deal with the problem of embedding property graphs in vector spaces. Throughout this paper we will use the term embedding as an operation that allows to consider a mathematical structure, X, inside another structure Y , through a function, f : X → Y . We are interested on embeddings capable of capturing, within the characteristics of a vector space (distance, linearity, clustering, etc.), the interesting features of the graph. For example, it would be interesting to get embeddings that, when projecting the nodes of the graph into points of a vector space, keep edges with the same type of the graph into the same vectors. In this way, we can interpret that the semantic associated to the relation has been captured by the embedding. Another option is to check if the embedding verifies clustering properties with respect to the types of nodes, types of edges, properties, or some of the metrics that can be measured on the graph. Subsequently, we will use these good embedding features to try to obtain prediction / classification / discovery tools on the original graph. This paper is structured as follows: we will start by giving some preliminary definitions necessary for the presentation of our proposal and a brief introduction to the use of artificial neural networks as encoding machines. After this review, we will present our embedding proposal based on neural encoders, and we will verify if the topological and semantic characteristics of the original graph have been maintained in the new representation. After evaluating the properties of the new representation, it will be used to carry out machine learning and discovery tasks on real databases. Finally, we will present some conclusions and future work proposals that have arisen during the implementation of this work.