Heterogeneity-Aware Continuous Graph Neural Diffusion Network for Unsupervised Fake News Detection (old)

This record provides the source code and data-access metadata for HCGDN, the method presented in “Heterogeneity-Aware Continuous Graph Neural Diffusion Network for Unsupervised Fake News Detection.” HCGDN is an unsupervised graph representation learning framework for fake news detection. It models heterogeneous propagation structures through two complementary components: a global graph neural diffusion process for capturing long-range propagation dependencies and a local semantics-driven diffusion process for representing intensive substructures. The release contains the model implementation, graph-construction and dataset-loading code, experiment configurations, evaluation procedures, and a locked dependency specification. The evaluation code uses stratified ten-fold MLP classification to assess the learned graph representations. The experiments reported in the associated manuscript use PolitiFact and GossipCop from FakeNewsNet, together with the Weibo benchmark described by Ma et al. The original news, user, comment, and social-media records are not redistributed in this record. Derived graph arrays and precomputed BERT feature matrices are also not included because they originate from third-party benchmark data and remain subject to the applicable upstream terms. PolitiFact and GossipCop should be obtained through the FakeNewsNet project: https://github.com/KaiDMML/FakeNewsNet The Weibo benchmark should be obtained from its original providers using the access and author-contact information associated with: https://doi.org/10.1145/3391250 After obtaining authorized access, users must prepare the local input files described in DATA_AVAILABILITY.md. README.md provides the software environment, directory structure, configuration details, and execution command. This is HCGDN software release version 1.0.0. The source code and documentation are distributed under the MIT License. The MIT License does not apply to FakeNewsNet, Weibo, or any original or derived dataset files.本数据集提供了发表于《面向无监督假新闻检测的异质性感知连续图神经扩散网络》（Heterogeneity-Aware Continuous Graph Neural Diffusion Network for Unsupervised Fake News Detection）的HCGDN方法的源代码与数据访问元数据。 HCGDN是一款面向假新闻检测的无监督图表示学习框架。它通过两个互补组件对异质传播结构进行建模：一是用于捕捉长距离传播依赖的全局图神经扩散过程，二是用于表征密集子结构的局部语义驱动扩散过程。 本次发布包含模型实现代码、图构建与数据集加载代码、实验配置、评估流程以及锁定的依赖项规范。 评估代码采用分层十折多层感知机（Multi-Layer Perceptron，MLP）分类方法对学习得到的图表示进行评估。 相关论文中报道的实验采用了来自FakeNewsNet的PolitiFact与GossipCop数据集，以及Ma等人提出的微博（Weibo）基准数据集。本数据集未重新分发原始新闻、用户、评论以及社交媒体记录。由于派生图数组与预计算的BERT（Bidirectional Encoder Representations from Transformers）特征矩阵均源自第三方基准数据，仍需遵循相应的上游使用条款，故未随本数据集一同发布。 PolitiFact与GossipCop数据集需通过FakeNewsNet项目获取：https://github.com/KaiDMML/FakeNewsNet Weibo基准数据集需通过其原始提供方，结合该数据集关联的访问方式与作者联系方式获取，相关链接为：https://doi.org/10.1145/3391250 在获得授权访问权限后，用户需按照DATA_AVAILABILITY.md中的说明准备本地输入文件。README.md文件提供了软件环境、目录结构、配置细节以及执行命令的相关信息。 本软件发布版本为HCGDN 1.0.0。源代码与文档采用MIT许可证进行分发，但MIT许可证不适用于FakeNewsNet、Weibo以及任何原始或派生数据集文件。