Dissertation published

Author

Max Berrendorf

Abstract

As the digitalization of private, commercial, and public sectors advances rapidly, an increasing amount of data is becoming available. In order to gain insights or knowledge from these enormous amounts of raw data, a deep analysis is essential. The immense volume requires highly automated processes with minimal manual interaction. In recent years, machine learning methods have taken on a central role in this task. In addition to the individual data points, their interrelationships often play a decisive role, e.g. whether two patients are related to each other or whether they are treated by the same physician. Hence, relational learning is an important branch of research, which studies how to harness this explicitly available structural information between different data points. Recently, graph neural networks have gained importance. These can be considered an extension of convolutional neural networks from regular grids to general (irregular) graphs.

Knowledge graphs play an essential role in representing facts about entities in a machine-readable way. While great efforts are made to store as many facts as possible in these graphs, they often remain incomplete, i.e., true facts are missing. Manual verification and expansion of the graphs is becoming increasingly difficult due to the large volume of data and must therefore be assisted or substituted by automated procedures which predict missing facts. The field of knowledge graph completion can be roughly divided into two categories: Link Prediction and Entity Alignment. In Link Prediction, machine learning models are trained to predict unknown facts between entities based on the known facts. Entity Alignment aims at identifying shared entities between graphs in order to link several such knowledge graphs based on some provided seed alignment pairs.

In this thesis, we present important advances in the field of knowledge graph completion. For Entity Alignment, we show how to reduce the number of required seed alignments while maintaining performance by novel active learning techniques. We also discuss the power of textual features and show that graph-neural-network-based methods have difficulties with noisy alignment data. For Link Prediction, we demonstrate how to improve the prediction for unknown entities at training time by exploiting additional metadata on individual statements, often available in modern graphs. Supported with results from a large-scale experimental study, we present an analysis of the effect of individual components of machine learning models, e.g., the interaction function or loss criterion, on the task of link prediction. We also introduce a software library that simplifies the implementation and study of such components and makes them accessible to a wide research community, ranging from relational learning researchers to applied fields, such as life sciences. Finally, we propose a novel metric for evaluating ranking results, as used for both completion tasks. It allows for easier interpretation and comparison, especially in cases with different numbers of ranking candidates, as encountered in the de-facto standard evaluation protocols for both tasks.

PDF available at edoc.ub.uni-muenchen.de.