Collaborative Research: SHF: Small: LEGAS: Learning Evolving Graphs At Scale

图神经网络是一类新兴的基于图的深度学习模型，在推荐系统、药物发现、社会网络分析、代码漏洞检测等领域有着广泛的应用。然而，由于GNN的计算涉及复杂的矩阵和向量运算，因此计算效率较低。此外，当应用于动态变化的图表时，效率问题会加剧。该项目率先致力于为静态和动态图形开发高效的GNN算法和计算系统，以利用世界级的图形处理单元(GPU)计算设施。该项目有助于国家对机器学习和计算系统专业人员日益增长的需求。这个项目产生了一个高性能的软件库，作为来自学术界、国家实验室和工业界的科学和工程从业者的基础工具。此外，教育部门还努力将研究成果纳入研究生和本科生课程开发。开展了外联和教育活动，以促进K-12、本科生、女性和代表性不足的少数群体的参与。这个项目的总体目标是通过静态和动态图形的算法和系统协同设计来设计一个高效的GNN框架。为此，本项目设计了三个协同研究推力。具体地说，推力1通过设计新的GNN算法从算法层面提高效率，这些算法是高效的，允许保留整个图，并提供收敛保证。推力2通过在GPU上设计高效的计算技术来提高系统性能，并具有高效的工作负载调度和减少的同步开销。此外，推力3结合了推力1和推力2中的技术，并设计了新的策略来解决动态图形中独特的算法和计算挑战。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Graph Neural Networks (GNNs) are an emerging class of deep learning models on graphs, with many successful applications, such as, recommendation systems, drug discovery, social network analysis, and code vulnerability detection. However, the computation for GNNs faces a low efficiency problem as they involve complex matrix and vector operations. Further, when applied to graphs that are dynamically changing, the efficiency issue exacerbates. This project pioneers the effort of developing efficient GNN algorithms and computation systems for both static and dynamic graphs that can take advantage of world-class Graphics Processing Unit (GPU) computing facilities. This project contributes to the growing national need for professionals in machine learning and computation systems. This project produces a high-performance software library that serves as a foundational tool for fellow science and engineering practitioners from academia, national laboratories, and industry. Additionally, educational efforts are made to integrate the research findings into graduate and undergraduate curriculum development. Outreach and educational activities are conducted to promote the participation of K-12, undergraduate, female, and underrepresented minorities. The overarching goal of this project is to design an efficient GNN framework via algorithm and system co-design for both static and dynamic graphs. Towards that, this project designs three synergistic research thrusts. Specifically, Thrust 1 improves the efficiency from the algorithm level by designing novel GNN algorithms that are efficient, allow the entire graph to be retained, and offer convergence guarantees. Thrust 2 advances the system performance by designing efficient computation techniques on GPUs with efficient workload scheduling and reduced synchronization overhead. In addition, Thrust 3 incorporates the techniques in Thrusts 1 and 2, and designs novel strategies to address the unique algorithm and computation challenges in dynamic graphs.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.