Predicting instance type assertions in knowledge graphs using stochastic neural networks

Instance type information is particularly relevant to perform reasoning and obtain further information about entities in knowledge graphs (KGs). However, during automated or pay-as-you-go KG construction processes, instance types might be incomplete or missing in some entities. Previous work focused mostly on representing entities and relations as embeddings based on the statements in the KG. While the computed embeddings encode semantic descriptions and preserve the relationship between the entities, the focus of these methods is often not on predicting schema knowledge, but on predicting missing statements between instances for completing the KG. To fill this gap, we propose an approach that first learns a KG representation suitable for predicting instance type assertions. Then, our solution implements a neural network architecture to predict instance types based on the learned representation. Results show that our representations of entities are much more separable with respect to their associations with classes in the KG, compared to existing methods. For this reason, the performance of predicting instance types on a large number of KGs, in particular on cross-domain KGs with a high variety of classes, is significantly better in terms of F1-score than previous work