FlowBanzhaf/FlowBanzhaf

FlowBanzhaf是一个大规模图基准数据集，包含约2000万个有向流网络实例，旨在系统评估图神经网络在受控分布偏移下的泛化能力。每个实例是一个有向流网络，其中代理控制边的子集；每个图的标签是归一化Banzhaf权力指数向量——衡量每个代理对最大流的边际贡献。数据集涵盖四种图结构家族：Erdős–Rényi（ER）、Barabási–Albert（BA）、Watts–Strogatz（WS）以及来自九个SNAP网络的真实世界子图，每个家族内系统变化图大小、边密度和代理数量。这使得它适用于研究参数偏移（固定生成过程内的变化）和结构偏移（跨图家族的迁移）。回归目标——归一化Banzhaf指数——精确计算是指数级的，需要全局路径推理，因此成为对消息传递GNN的原则性压力测试，因为其局部聚合限制了长程依赖捕获。数据集的迷你版本（所有家族中每个配置100个图）也可用于快速实验。FlowBanzhaf is a large-scale graph benchmark of approximately 20 million directed flow network instances, designed to support systematic evaluation of graph neural network (GNN) generalization under controlled distribution shift. Each instance is a directed flow network in which agents control subsets of edges; the label for each graph is the vector of normalized Banzhaf power indices — a measure of each agents marginal contribution to the maximum flow. The dataset spans four structural graph families — Erdős–Rényi (ER), Barabási–Albert (BA), Watts–Strogatz (WS), and real-world subgraphs from nine SNAP networks — with systematic variation of graph size, edge density, and agent count within each family. This makes it suitable for studying both parametric shift (variation within a fixed generative process) and structural shift (transfer across graph families). The regression target — the normalized Banzhaf index — is exponential to compute exactly and requires global path reasoning, making it a principled stress test for message-passing GNNs whose local aggregation limits long-range dependency capture. A mini version of the dataset (100 graphs per configuration across all families) is also available for quick experimentation.