Fake news detection (FND) is an application-oriented hotspot in the field of artificial intelligence, whose task is to make neural networks to judge the authenticity of given news, and the challenge it faces is how to train neural networks effectively. Currently, state-of-the-art approaches typically employ the methods based on graph convolutional neural networks (GCNs) to extract features of news dissemination. However, these methods cannot effectively represent temporal features and cannot handle the problem of imbalanced positive and negative samples. This motivates us to investigate the impact of temporal information on fake news detection and the impact of sample balance on model training. To this end, we propose a social media ake ews etection model based on idirectional emporal-delay raph onvolution etwork (). In , we extend unidirectional graphs to bidirectional graphs and design bidirectional temporal-delay graph convolutional networks to effectively represent graph-structured data. We further design heuristic graph-structured data enhancement strategies to fully leverage information. Moreover, we introduce a graph contrastive learning method, which improves the model performance by computing the mutual information between positive and negative samples. We have conducted experimental researches on two publicly available real-world datasets. The experimental results show that compared with the current state-of-the-art methods, our model has achieved an average improvement of 2.2% in detection accuracy and 1.9% in F1-score.

中文翻译：

---

用于检测假新闻的双向时延图卷积网络

假新闻检测（FND）是人工智能领域面向应用的热点，其任务是让神经网络判断给定新闻的真实性，其面临的挑战是如何有效地训练神经网络。目前，最先进的方法通常采用基于图卷积神经网络（GCN）的方法来提取新闻传播的特征。然而，这些方法不能有效地表示时间特征，也不能处理正负样本不平衡的问题。这促使我们研究时间信息对假新闻检测的影响以及样本平衡对模型训练的影响。为此，我们提出了一种基于单向时延拉夫 onvolution etwork 的社交媒体 ake ews 检测模型（）。在 中，我们将单向图扩展到双向图，并设计双向时延图卷积网络来有效地表示图结构数据。我们进一步设计启发式图结构数据增强策略以充分利用信息。此外，我们引入了一种图对比学习方法，该方法通过计算正样本和负样本之间的互信息来提高模型性能。我们对两个公开的真实世界数据集进行了实验研究。实验结果表明，与当前最先进的方法相比，我们的模型在检测精度上平均提高了 2.2％，在 F1 分数上平均提高了 1.9％。