Temporal relational reasoning with neural networks

Abstract

The study of relationships between objects and their properties is a common pattern in many areas of science.Studies on developmental disorders, gene expression analysis, and rumor propagation in social networks are some of the many examples where the analysis of graph data is fundamental to comprehending the phenomena in question.However, manually extracting and reasoning about this information quickly become a daunting undertaking once the number of connections and entities most of these phenomena involve is taken into account.Artificial neural networks (NNs) have recently been a central piece of many state-of-the-art machine learning methods, aimed exactly at dealing with large volumes of data.In particular, there has been an increasing interest in applying them to graph domains.Still, one important fact mostly unaccounted for is that almost all large real-world graphs evolve over time.In this research, we aim to identify relevant phenomena that can be represented as dynamic graphs and, in turn, to design neural network architectures that can be used to solve problems involving those phenomena.We adopt a methodology consisting of three stages for this.First, we establish baselines and comparisons for existing NNs designed for graph domains, since most of them were not directly compared to each other.We also take this opportunity to highlight which of those NNs also have a potential to be extended to dynamic graph domains.In the second stage, we extend those baselines by changing the domain to dynamic graphs.The third stage is concerned with tangible learning scenarios.For example, we expect to employ the developed methods in soccer match analysis.The reasoning for this is that many of the common learning tasks with graph data can be employed to soccer match analysis by modeling players as nodes (entities) and their interactions as edges (relationships).As a complex system of interactions, we expect that this leads to challenging problems not explored before.We expect, as a result of this research, NNs that successfully learn on dynamic graph domains.This will lead to contributions not only to machine learning, but also to other areas of science where we can represent a phenomenon with dynamic graphs. (AU)