The upsurges of location-based social media and geo-location based online social networks have accelerated the spread of rumors, making rumor detection an increasingly important challenge. To tackle this challenge, a variety of conventional and deep learning based approaches are developed to detect rumors based on geo-textual data. This data includes textual, spatial, and temporal information generated from a variety of location-aware social media services. However, the effectiveness of early-stage detection methods based on geo-textual data is limited due to additional information deficiencies, such as the lack of comments and retweets. To address this limitation, we propose a model named Dual-Level Network for Rumor Detection (DLRD). The DLRD model extracts both post-level features and topic-level features. A topic-level network is applied to capture the features of geo-textual data with similar expressions and meanings, such as those that share the same geographic location and time point, indicating that they have a high probability of sharing the same truth value. Specifically, we first leverage the topic modeling method to analyze text content including spatio-temporal information, divide source posts of all events into disjoint topic groups, and then construct a topic-post graph for each group. In the DLRD model, two graph convolutional layers and two full connection layers are employed to learn topic-level and post-level features, respectively. We conduct extensive experiments to compare our model against existing baselines on two public real-world datasets. The experimental results demonstrate that the DLRD model achieves state-of-the-art performance over existing baselines.

基于位置的社交媒体和基于地理位置的在线社交网络的热潮加速了谣言的传播，使得谣言检测成为越来越重要的挑战。为了应对这一挑战，开发了各种基于传统和深度学习的方法来基于地理文本数据检测谣言。这些数据包括从各种位置感知社交媒体服务生成的文本、空间和时间信息。然而，由于缺乏评论和转发等额外信息，基于地理文本数据的早期检测方法的有效性受到限制。为了解决这个限制，我们提出了一种名为双层网络谣言检测（DLRD）的模型。DLRD模型同时提取帖子级特征和主题级特征。应用主题级网络来捕获具有相似表达和含义的地理文本数据的特征，例如具有相同地理位置和时间点的地理文本数据，表明它们很有可能共享相同的真值。具体来说，我们首先利用主题建模方法来分析包括时空信息的文本内容，将所有事件的源帖子划分为不相交的主题组，然后为每个组构建主题帖子图。在DLRD模型中，两个图卷积层和两个全连接层分别用于学习主题级和后级特征。我们进行了广泛的实验，将我们的模型与两个公共现实数据集的现有基线进行比较。