Logo Logo
FAQ
Contact
Switch language to German
Representation learning on relational data
Representation learning on relational data
Humans utilize information about relationships or interactions between objects for orientation in various situations. For example, we trust our friend circle recommendations, become friends with the people we already have shared friends with, or adapt opinions as a result of interactions with other people. In many Machine Learning applications, we also know about relationships, which bear essential information for the use-case. Recommendations in social media, scene understanding in computer vision, or traffic prediction are few examples where relationships play a crucial role in the application. In this thesis, we introduce methods taking relationships into account and demonstrate their benefits for various problems. A large number of problems, where relationship information plays a central role, can be approached by modeling data by a graph structure and by task formulation as a prediction problem on the graph. In the first part of the thesis, we tackle the problem of node classification from various directions. We start with unsupervised learning approaches, which differ by assumptions they make about the relationship's meaning in the graph. For some applications such as social networks, it is a feasible assumption that densely connected nodes are similar. On the other hand, if we want to predict passenger traffic for the airport based on its flight connections, similar nodes are not necessarily positioned close to each other in the graph and more likely have comparable neighborhood patterns. Furthermore, we introduce novel methods for classification and regression in a semi-supervised setting, where labels of interest are known for a fraction of nodes. We use the known prediction targets and information about how nodes connect to learn the relationships' meaning and their effect on the final prediction. In the second part of the thesis, we deal with the problem of graph matching. Our first use-case is the alignment of different geographical maps, where the focus lies on the real-life setting. We introduce a robust method that can learn to ignore the noise in the data. Next, our focus moves to the field of Entity Alignment in different Knowledge Graphs. We analyze the process of manual data annotation and propose a setting and algorithms to accelerate this labor-intensive process. Furthermore, we point to the several shortcomings in the empirical evaluations, make several suggestions on how to improve it, and extensively analyze existing approaches for the task. The next part of the thesis is dedicated to the research direction dealing with automatic extraction and search of arguments, known as Argument Mining. We propose a novel approach for identifying arguments and demonstrate how it can make use of relational information. We apply our method to identify arguments in peer-reviews for scientific publications and show that arguments are essential for the decision process. Furthermore, we address the problem of argument search and introduce a novel approach that retrieves relevant and original arguments for the user's queries. Finally, we propose an approach for subspace clustering, which can deal with large datasets and assign new objects to the found clusters. Our method learns the relationships between objects and performs the clustering on the resulting graph.
representation learning, relational learning
Faerman, Evgeniy
2021
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Faerman, Evgeniy (2021): Representation learning on relational data. Dissertation, LMU München: Faculty of Mathematics, Computer Science and Statistics
[thumbnail of Faerman_Evgeniy.pdf]
Preview
PDF
Faerman_Evgeniy.pdf

898kB

Abstract

Humans utilize information about relationships or interactions between objects for orientation in various situations. For example, we trust our friend circle recommendations, become friends with the people we already have shared friends with, or adapt opinions as a result of interactions with other people. In many Machine Learning applications, we also know about relationships, which bear essential information for the use-case. Recommendations in social media, scene understanding in computer vision, or traffic prediction are few examples where relationships play a crucial role in the application. In this thesis, we introduce methods taking relationships into account and demonstrate their benefits for various problems. A large number of problems, where relationship information plays a central role, can be approached by modeling data by a graph structure and by task formulation as a prediction problem on the graph. In the first part of the thesis, we tackle the problem of node classification from various directions. We start with unsupervised learning approaches, which differ by assumptions they make about the relationship's meaning in the graph. For some applications such as social networks, it is a feasible assumption that densely connected nodes are similar. On the other hand, if we want to predict passenger traffic for the airport based on its flight connections, similar nodes are not necessarily positioned close to each other in the graph and more likely have comparable neighborhood patterns. Furthermore, we introduce novel methods for classification and regression in a semi-supervised setting, where labels of interest are known for a fraction of nodes. We use the known prediction targets and information about how nodes connect to learn the relationships' meaning and their effect on the final prediction. In the second part of the thesis, we deal with the problem of graph matching. Our first use-case is the alignment of different geographical maps, where the focus lies on the real-life setting. We introduce a robust method that can learn to ignore the noise in the data. Next, our focus moves to the field of Entity Alignment in different Knowledge Graphs. We analyze the process of manual data annotation and propose a setting and algorithms to accelerate this labor-intensive process. Furthermore, we point to the several shortcomings in the empirical evaluations, make several suggestions on how to improve it, and extensively analyze existing approaches for the task. The next part of the thesis is dedicated to the research direction dealing with automatic extraction and search of arguments, known as Argument Mining. We propose a novel approach for identifying arguments and demonstrate how it can make use of relational information. We apply our method to identify arguments in peer-reviews for scientific publications and show that arguments are essential for the decision process. Furthermore, we address the problem of argument search and introduce a novel approach that retrieves relevant and original arguments for the user's queries. Finally, we propose an approach for subspace clustering, which can deal with large datasets and assign new objects to the found clusters. Our method learns the relationships between objects and performs the clustering on the resulting graph.