52442 research outputs found
Sort by
Political trust during the COVID-19 pandemic in Germany. Insights from a survey experiment
No full text
Neural methods for link prediction in knowledge graphs
Get PDFIntegrating neural models with multi-relational data poses a significant challenge, mainly due to the complexity and highly interconnected nature of such data, which is prevalent in domains such as social networks, knowledge graphs, and biomedical databases. In this context, this thesis explores three key challenges: effectively capturing multi-relational data for efficient reasoning and integration into downstream applications, integrating structural and textual information, and managing the evolving nature of multi-relational graphs. Knowledge graph embeddings (KGE) offer a promising approach, learning low-dimensional representations for entities and relations to enhance tasks such as recommendation, question answering, and visual relationship detection. However, scalability and quality assurance in KGE models remain issues, especially with large-scale graphs. Recent frameworks address scalability through parallelization techniques, but their impact on model quality requires further study. While extensive hyperparameter optimization can considerably improve the resulting embedding quality, current optimization methods do not scale to large-scale graphs. Moreover, integrating textual information with structural data in a simple, scalable, and versatile manner remains an open problem.
Addressing these challenges, this thesis makes several key contributions. First, it evaluates and improves the efficiency and effectiveness of parallel training techniques for KGE models. By re-implementing and extensively testing various parallelization methods, the study identifies optimal techniques for large-scale KGE training, demonstrating significant speedups while maintaining model quality. Second, it proposes GraSH, a scalable hyperparameter optimization algorithm based on successive halving, achieving state-of-the-art results on large-scale knowledge graphs with minimal search budgets. Third, the thesis introduces KGT5-context, a novel approach that integrates structural and textual information by posing link prediction as a sequence-to-sequence task, achieving or surpassing state-of-the-art results in link prediction and downstream tasks. Finally, it presents the Wikidata5M-SI benchmark for evaluating model performance on dynamic graphs with emerging entities in a realistic, large-scale setting
