Predicting the co-evolution of event and Knowledge Graphs

Knowledge graphs have evolved as flexible and powerful means for representing general world knowledge. Typical examples are DBpedia, Yago, or the Google Knowledge Graph, which all started off by representing information derived from Wikipedia and were then greatly expanded. In this paper we use the concept of a knowledge graph to present information about specific classes of entities, such as patients or users. The knowledge graph represents all that is known about the entities and their relationships and the goal is to integrate and exploit that information for prediction and decision support. In previous papers it was shown that embedding learning, a.k.a. representation learning, is capable of modelling large-scale semantic knowledge graphs, by exploiting information that describes the context of an entity in the knowledge graph. In Machine Learning we often map the knowledge graph to a tensor representation. Then we learn the latent representations of the entities that compose the tensor and use them to predict unobserved facts. However knowledge graphs represent the current status of the world and therefore they lack of a temporal dimension, which means we can only use them to predict facts about the present moment. In this paper we introduce an additional set of tensors that contain temporal information. Each of this event tensors contains all the events that occurred on a particular time step. Our goal will be to predict the events that will happen in future time steps, using for that task both dynamic information from the previous event tensors and static information that is stored in the knowledge graph. Therefore, this architecture allows us to fuse static and dynamic information to predict future events. We present experiments showing how this model performs well in multiple scenarios: medical data, a recommendation engine and sensor data.