Cora, Citeseer, CoAuthorCS, Polblogs and SBM

1.Cora dataset is derived from a multi-group citation network, and the two-group subgraphs are selected for tasks such as graph neural network node classification. The dataset contains sparse Bag-of-Words feature vectors as node attributes, and the labels are mostly academic paper topic categories or fields. This subgraph focuses on the influence of graph structure and node characteristics on model prediction, which provides a reliable experimental benchmark for the research of multi-step adversarial attacks and defense strategies.

Number of nodes: (example) 652
Edges: (example) 2350
Node feature dimension: 1433
Applicable tasks: node classification, adversarial attack, graph representation learning, etc

2.Citeseer is also derived from multi-group citation networks and is similar to Cora, but differs in node distribution and feature dimensions. In this double-group subgraph, the node attributes also use sparse bag-of-words feature vectors, and the labels are mostly research topics or directions of academic papers. Because the graph structure is relatively complex, and the node feature dimensions and the number of categories are different from Cora, this dataset is often used to compare and verify the generalization and robustness of graph neural network models.

Number of nodes: (example) 852
Edges: (example) 3170
Node feature dimension: 3703
Applicable tasks: node classification, adversarial attack, citation network analysis, etc

3.CoAuthorCS comes from the two-group subgraph of the multi-group cooperation network, and each node represents the presence of keywords by a binary feature vector, which is suitable for studying the task of clustering or classification based on the presence or not of attributes. This dataset can highlight the association between node characteristics and cooperation relationships in academic networks, and provide experimental scenarios with more binary attribute characteristics for multi-step adversarial attack research.

Number of nodes: (example) 836
Edges: (example) 2270
Node feature form: binary keyword vector
Applicable tasks: node classification, cooperative relationship analysis, adversarial attack, etc

4.Polblogs is a real-world dataset that reflects a network of political blogs, with node labels corresponding to the political orientation of the blogs (e.g., liberal vs. conservative). The network structure of the dataset is usually large, and the edges represent the reference or link relationships between blogs. It is often used to analyze community division, public opinion diffusion, adversarial attacks, and so on. By treating node labels as binary classes (Liberal vs. Conservative), researchers can test the effectiveness of adversarial attacks and defense mechanisms in complex community structures.

Number of nodes: (example) 1222
Edges: (example) 16714
Tag type: Liberal/Conservative
Applicable tasks: node classification, community division, polarization research, adversarial attack, etc

5.Stochastic Block Model (SBM) is a commonly used stochastic graph model to simulate network data with community structure or block structure. The data set can be generated by random generation mechanism (such as setting the number of communities, edge probability, etc.), and the node label is often determined by the community it belongs to. SBM is often used to study community detection, group behavior simulation, and robustness under adversarial attacks, because of its high controllability and the ability to adjust the network size and structure according to requirements.

Number of nodes: (example) 1490
Edges: (example) 13790
Label type: Community division based on synthesis
Applicable tasks: community detection, random graph model research, adversarial attack simulation, etc

1. Cora数据集源自多组引用网络，选取其二组子图用于图神经网络节点分类等任务。该数据集包含稀疏词袋（Bag-of-Words）特征向量作为节点属性，标签多为学术论文的主题类别或研究领域。该子图聚焦于图结构与节点特征对模型预测的影响，为多步对抗攻击与防御策略的研究提供了可靠的实验基准。
节点数：（示例）652
边数：（示例）2350
节点特征维度：1433
适用任务：节点分类、对抗攻击、图表示学习等

2. Citeseer同样源自多组引用网络，与Cora数据集类似，但在节点分布与特征维度上存在差异。在该二组子图中，节点属性同样采用稀疏词袋（Bag-of-Words）特征向量，标签多为学术论文的研究主题或方向。由于其图结构相对复杂，且节点特征维度与类别数均与Cora不同，该数据集常被用于对比验证图神经网络模型的泛化性与鲁棒性。
节点数：（示例）852
边数：（示例）3170
节点特征维度：3703
适用任务：节点分类、对抗攻击、引用网络分析等

3. CoAuthorCS源自多组合作网络的二组子图，每个节点通过二元特征向量表示关键词的存在性，适用于基于属性存在与否的聚类或分类任务。该数据集能够凸显学术网络中节点特征与合作关系之间的关联，为多步对抗攻击研究提供了具备更多二元属性特征的实验场景。
节点数：（示例）836
边数：（示例）2270
节点特征形式：二元关键词向量
适用任务：节点分类、合作关系分析、对抗攻击等

4. Polblogs是反映政治博客网络的真实数据集，节点标签对应博客的政治倾向（如自由派 vs 保守派）。该数据集的网络结构通常规模较大，边代表博客之间的引用或链接关系，常被用于社区划分、舆论传播、对抗攻击等研究。通过将节点标签视为二元类别（自由派/保守派），研究人员可测试对抗攻击与防御机制在复杂社区结构中的有效性。
节点数：（示例）1222
边数：（示例）16714
标签类型：自由派/保守派
适用任务：节点分类、社区划分、极化研究、对抗攻击等

5. 随机块模型（Stochastic Block Model，SBM）是一种常用于模拟具备社区结构或块结构的网络数据的随机图模型。该数据集可通过随机生成机制构建（如设置社区数量、边概率等参数），节点标签通常由其所属社区决定。由于其具备高度可控性，可根据需求调整网络规模与结构，因此常被用于社区检测、群体行为模拟以及对抗攻击下的鲁棒性研究。
节点数：（示例）1490
边数：（示例）13790
标签类型：基于合成的社区划分
适用任务：社区检测、随机图模型研究、对抗攻击模拟等

搜集汇总

背景与挑战

背景概述

该数据集是一个包含Cora、Citeseer、CoAuthorCS、Polblogs和SBM五个图数据集的集合，涵盖引用网络、合作网络、政治博客网络和合成随机网络等多种类型。这些数据集提供稀疏词袋特征向量、二进制特征向量和节点标签，适用于图神经网络节点分类、多步对抗攻击、社区检测等研究任务，旨在为图相关实验提供统一的基准测试平台。