FAGRE: A Fast Graph Neural Network Redundancy Elimination System

Graph Neural Networks (GNNs) are pivotal in analyzing complex relationships in recommendation systems and social networks. However, existing redundancy elimination methods for optimizing GNNs often face challenges with large-scale graphs, resulting in significant overhead. In this work, we introduce the Fast GNN Redundancy Elimination System (FAGRE), which includes a Pruning Optimization Module to minimize iteration scale and an Index-Multiplication Module to enhance computational speed. These modules collectively reduce time complexity and operational overhead while ensuring effective redundancy elimination, thereby improving the processing speed of redundancy elimination tasks on large-scale graphs. FAGRE can be seamlessly integrated into existing workflows by replacing specific components. Experimental results from five large-scale datasets demonstrate that the two SOTA redundancy elimination methods integrated with FAGRE achieve a speedup of up to 3.6× in total GNN inference tasks, while reducing the computational overhead of the redundancy elimination system by over 99%. Our code is available at https://github.com/Anonymous-615/Fa-GRE.