Task-Oriented Embedding for Node Heaviness Prediction in Graphs

In recent years, graph analysis has gained significant traction due to the widespread presence of networks across various domains, including biology, linguistics, and social sciences. This thesis explores the intricate relationship between node embedding techniques and the accuracy of motif estimation in complex networks, highlighting the critical role that embedding methods play in understanding network structures. We introduce the concept of Node Heaviness, which quantifies a node's involvement in motifs, and propose a novel node embedding technique specifically designed for predicting node heaviness. Through comprehensive experiments, we evaluate several embedding methods, including matrix factorization approaches, random walks, and deep learning techniques, with the goal of enhancing the accuracy of node heaviness predictions by capturing structural characteristics of nodes. Our analysis reveals varying effectiveness based on dataset characteristics and the specific nature of the graphs. While deep learning methods generally exhibit superior performance, they often struggle with dense graphs, where the complexity of the network structure can hinder their effectiveness. In contrast, our custom node embedding technique tailored for this task demonstrates adequate performance, showcasing significant potential for improvement. By analyzing the strengths and weaknesses of different embedding strategies, this thesis contributes valuable insights to the field of network analysis. The findings underscore the importance of selecting appropriate embedding techniques for specific graph characteristics.