Sadeghi, Afshin: Multiple Objective Learning for Effective Knowledge Graph Embedding. - Bonn, 2023. - Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn.
Online-Ausgabe in bonndoc:
author = {{Afshin Sadeghi}},
title = {Multiple Objective Learning for Effective Knowledge Graph Embedding},
school = {Rheinische Friedrich-Wilhelms-Universität Bonn},
year = 2023,
month = mar,

note = {Over the past decade, knowledge graphs (KGs) have become popular for capturing structured domain knowledge. Knowledge graphs particularly allow the effortless integration of heterogeneous data into a coherent model. Besides applications in data integration, KGs are at the center of many artificial intelligence studies as an expressive data model for representation learning purposes. Knowledge graph embedding (KGE) methods produce a latent representation of KGs entities showing applicability potential for solving downstream tasks, including link prediction and node classification. KGE also supports Named Entity Resolution in NLP tasks and is applied in Question Answering Systems.
Existing KGE models have achieved excellent results over simple knowledge graphs, where they contain only a few relation patterns that leak into each other. Besides, in simple knowledge graphs, the amount of entities with similar neighbors is lower, and the structure of the subgraphs is unique, so their entities are more easily distinguished.
This work dives more into the study of KGE for complex knowledge graphs. In such KGs, distinct relation patterns show up significantly more, and similar substructures repeat over the network on a larger scale. Therefore, recognizing unique entities with limited knowledge about the direct neighbors and the limited recognition of relation patterns is remarkably more difficult.
Complex knowledge graph embedding provides several challenges, such as understanding learning distinct relation patterns and graphical features of nodes. The lack of suitable datasets that emulate the difficulty of more complex knowledge graphs further adds to research gaps. Hence, in this thesis, we focus on the research objective of laying the foundations for the advancement of the state-of-the-art to better embed complex knowledge graphs by providing techniques to solve various challenges and resources to fill the research gaps.
First, to effectively target the complex KGE challenge, we propose a multi-objective method that allows learning several relation patterns of knowledge graphs. Multiple Distance Embedding (MDE) generalizes over several distance-based models and proposes combined learning objectives that extract more knowledge from KGs as the base training data. We demonstrate that MDE allows modeling relations with (anti)symmetry, inversion, and composition patterns. Our empirical investigation shows the on-par and better performance of MDE relative to the state-of-the-art methods in the link prediction task and its effectiveness in learning individual relation patterns.
We then propose a novel KGE method named Graph Feature Attentive Neural Network (GFA-NN) that computes graphical features of entities. This method addresses the limitation of embedding models that consider only the local graph structure related to an entity and have difficulty distinguishing similar graph substructures.
Consequently, the resulting embeddings of GFA-NN are attentive to two types of global network features. First, nodes’ relative centrality is based on the observation that some entities are more "prominent" than others. The second is the relative position of entities in the graph. GFA-NN computes several centrality values per entity, generates a random set of reference entities, and computes a given entity’s shortest path to each entity in the reference set. It then learns this information by optimizing the objectives specified on each feature. We investigate GFA-NN on several link prediction benchmarks in the inductive and transductive settings. We demonstrate that GFA-NN improves the state-of-the-art records of KGE solutions, especially on large-scale and more complex knowledge graphs.
We next construct 96 datasets replicating 16 different relation pattern circumstances and four different inductive and transductive test settings. This effort addresses the gap of missing a leak-free link prediction benchmarks. Our analysis of embedding models over these datasets provides a better insight into the suitable parameters for each situation, optimizing the KG-embedding-based systems.},

url = {}

The following license files are associated with this item: