Cost-Effective Machine Learning Techniques for Graphs

Graph-structured data is prevalent in various domains, including social, commercial, and molecular networks. To capture the underlying complexity of such data, graph neural networks (GNNs) have emerged as a powerful tool, delivering exceptional performance across tasks like node classification, link prediction, and node clustering. However, applying GNNs in large-scale industrial settings presents significant challenges, particularly when dealing with graphs with massive numbers of nodes and edges. Additionally, despite the high computational costs, GNNs often struggle to achieve consistently strong performance across all tasks and node cohorts, highlighting the need for more cost-effective approaches. This dissertation addresses these challenges by improving the cost-effectiveness of GNNs through four key dimensions: optimization, model design, task alignment, and data handling. In terms of optimization, we developed methods that substantially accelerate GNN training without compromising performance. From model, task, and data perspectives, we proposed frameworks that enable GNNs to perform effectively across multiple tasks and diverse node cohorts simultaneously. These contributions collectively enhance the scalability and practicality of GNNs, offering solutions to make them more efficient and applicable to real-world social network mining.