FairCAT-generated datasets for benchmarking fairness-aware GNNs

These synthetic datasets are created with an aim to be used for benchmarking fairness and performance of Graph Neural Networks (GNNs). The datasets are created specifically for the following experiments: balancing, correlations strength, data scaling, and synthetic German Credit, and synthetic Pokec_n. The environment for each experiment varies in only one variable, according to the experiment, while others stay fixed. Balancing experiment: the ratio of the sensitive groups is varied for each graph: Balanced (50/50 split), Mild Imbalance (70/30), and Strong Imbalance (90/10). Correlation strength experiment: graphs with varying strength of correlation between sensitive attribute and a non-sensitive attribute: Low (0.05), Medium (0.50), and High (0.95). Scaling experiment: graphs are increasing in size: Small (number of nodes=2^15), Medium (2^20), Large (2^23). German FairCAT: a synthetic dataset that tries to reproduce real-world German Credit dataset. It can be used to compare GNN predictions with learning on the original dataset. Pokec_n FairCAT: a synthetic dataset that tries to reproduce real-world Pokec_n dataset.本合成数据集系列旨在用于图神经网络（Graph Neural Networks，GNN）的公平性与性能基准测试。本数据集专为以下四类实验及两类复刻现实数据集的任务打造：平衡性实验、相关性强度实验、数据缩放实验，以及合成版德国信贷数据集与合成版Pokec_n数据集的构建与测试。每项实验仅对单一变量进行调整，其余变量均保持固定。 平衡性实验：每张图的敏感群体占比设置为不同水平，涵盖完全平衡（50:50划分）、轻度失衡（70:30）与重度失衡（90:10）三种场景。 相关性强度实验：构建敏感属性与非敏感属性间相关性强度各异的图结构，分为低相关性（0.05）、中相关性（0.50）与高相关性（0.95）三个档位。 数据缩放实验：图的规模依次递增，分为小规模（节点数=2^15）、中规模（2^20）与大规模（2^23）三个等级。 德国公平性基准数据集（German FairCAT）：旨在复刻现实世界中的德国信贷数据集，可用于对比图神经网络在该合成数据集与原始数据集上的预测效果差异。 Pokec_n公平性基准数据集（Pokec_n FairCAT）：旨在复刻现实世界中的Pokec_n数据集。