Unveiling Real-World Networks: Exploring Natural Language Models and Graph Neural Networks for Pattern Mining

Author:

Rao, Susie Xi [claim] ; id_orcid:0 000-0003-2379-1506 [claim]

Description:

This thesis explores real-world network analytics using a combination of natural language processing (NLP) methods and graph neural network (GNN) models. The study of real-world networks has gained significant attention due to the need to extract insights and uncover hidden patterns within complex interconnected data. My research aims to advance our understanding of real-world transaction and science networks and their underlying dynamics in various domains. The thesis addresses several research objectives and questions related to explainable anomaly detection in transaction networks in both online and offline settings, the large-scale hierarchical classification of research fields, and the quantification of interdisciplinarity in science networks. It involves the development of novel techniques and systems tailored to the challenges of real-world networks, including data collection, preprocessing, training and learning, model evaluation, and continuous learning. The contributions of this thesis can be summarized into two main aspects. First, in fraud detection in transaction networks, we have successfully developed systems of explainable solutions that can be deployed on billion-scale graphs and benchmarked our efforts against human annotators. This includes the development of explainable anomaly detection algorithms for transaction graphs and efficient fraud detection in dynamic graphs. Second, our work has led to the construction of the largest hierarchical classification system that encompasses all disciplines and scholarly publications. This system plays a crucial role within our comprehensive ecosystem of studying science networks, which comprises a hierarchical classification system for research fields, a scientific annotation and inference engine, and a framework for quantifying interdisciplinarity in science networks. These contributions improve our understanding of real-world transaction and science networks, by enabling efficient and accurate fraud detection systems, and providing insights into ...

Publisher:

ETH Zurich

Contributors:

Zhang, Ce ; Egger, Peter ; Schneider, Gerold

Year of Publication:

2023

Document Type:

info:eu-repo/semantics/doctoralThesis ; [Doctoral and postdoctoral thesis]

Language:

Subjects:

info:eu-repo/classification/ddc/004 ; Data processing ; computer science

DDC: