Temporal reasoning in knowledge graphs : Artificial intelligence systems for reasoning with time in Vadalog

Abstract

The rise of knowledge graphs has sparked great interest in providing scalable and efficient reasoning capabiliies for a variety of problems. A particularly prominent language supporting scalable reasoning techniques is Vadalog, which supports advanced reasoning capabilities such as existential quantification, recursion as well as aggregation, probabilistic reasoning, and various data sources and has demonstrated its value in numerous financial applications, including company control and golden power checks. In all of these applications, time is a critical dimension to gain a deeper understanding of the structural changes. However, so far, Vadalog is missing the support for dealing with temporal information, limiting its applicability in temporal contexts. The absence of such functionality is emphasized by the resurgence of temporal reasoning in the context of stream reasoning through DatalogMTL, an extension of Datalog with operators from the metric temporal logic. Yet, since DatalogMTL is a merely extension of Datalog, it lacks many of the capabilities necessary for knowledge graph reasoning. As a result, in this thesis, we conduct the first study on how to extend DatalogMTL towards its application in knowledge graph reasoning. For this purpose, we first study extensions of DatalogMTL, namely aggregation and existential quantification, which are fundamental to numerous data science workflows. In detail, we define formal syntax and semantics, explore different possibilities for aggregating along the timeline as well as examine a natural as well as a uniform semantic for existential quantification. Subsequently, we present a novel benchmark generator that is the first of its kind which is capable of supporting the generation of benchmarks for metric temporal logic, together with recursive queries, aggregation and existential quantification. This allows us to generate targeting instances for testing specific scenarios and edge cases. Afterwards, we augment Vadalog with the ability to reason with metric temporal logic providing a fully engineered reasoning architecture. We evaluate the system with benchmarks generated from our generator as well as from real-world instances. The results show that our system outperforms state-of-the art solutions in most of the scenarios. Finally, we discuss the usage of DatalogMTL as specification language of smart contracts, enabling the use of DatalogMTL for decentralized finance, an interesting domain for knowledge graph reasoning.