Abstract With the popularity of social networks, terrorist groups such as ISIS encouraged others to follow their activities, share their ideas, recruit fans, radicalize communities, and raise funds to support future attacks. This has led to the emergence of radicalized online accounts that belong to terrorists or their fans. Existing techniques for counter-terrorism investigations which aim to suspend such accounts are based on reports by users or syntactic-based sentiment analysis techniques, which are not accurate on short texts shared by terrorist such as tweets. This work proposed a feature augmentation approach to enrich the content of tweets before investigating them to discover the radicalized online contents. The augmented tweets are then used to classify accounts into Pro-ISIS or Anti-ISIS categories. We utilized topic modeling as a baseline method for feature augmentation. We studied the effects of utilizing tweets at different time intervals on the quality of the generated models that classify tweets and the corresponding accounts. We then introduced a novel feature augmentation approach that utilizes Neighborhood Overlap, a graph proximity technique that discovers terms having a strong relationship with the Pro-ISIS category. Terms extracted from tweets are represented as nodes in a graph, which is then partitioned into clusters containing different terms. Terms in strongly connected parts of each cluster are augmented to the original term vectors of the tweets based on the similarity between those terms and each tweet. We compared our approach with other baseline augmentation techniques such Term-to-Term correlation, Topic Modeling, and other existing techniques. Experimental results on a dataset containing Pro- and Anti-ISIS tweets show that our approach is quite promising to automate the identification of terrorist contents online. The results have shown that using graph proximity measures such as Neighborhood Overlap for term augmentation gains higher Precision, Recall, and F-measure than the typical approaches. In addition, we found that applying time-based analysis with term augmentation to identify radicalized accounts enhanced the Precision of the investigation process.

中文翻译：

---

一种用于识别 Twitter 上恐怖分子账户的图邻近特征增强方法

摘要 随着社交网络的流行，ISIS 等恐怖组织鼓励其他人关注他们的活动、分享他们的想法、招募粉丝、激进社区并筹集资金以支持未来的袭击。这导致了属于恐怖分子或其粉丝的激进在线帐户的出现。旨在暂停此类帐户的现有反恐调查技术基于用户的报告或基于句法的情感分析技术，这些技术对于恐怖分子共享的短文本（如推文）并不准确。这项工作提出了一种特征增强方法来丰富推文的内容，然后再调查它们以发现激进的在线内容。然后使用增强的推文将帐户分类为 Pro-ISIS 或 Anti-ISIS 类别。我们利用主题建模作为特征增强的基线方法。我们研究了在不同时间间隔使用推文对分类推文和相应帐户的生成模型质量的影响。然后，我们引入了一种新颖的特征增强方法，该方法利用邻域重叠，这是一种图形邻近技术，可发现与 Pro-ISIS 类别有密切关系的术语。从推文中提取的术语在图中表示为节点，然后将其划分为包含不同术语的集群。每个集群的强连接部分中的术语根据这些术语与每条推文之间的相似性被扩充为推文的原始术语向量。我们将我们的方法与其他基线增强技术进行了比较，例如术语到术语的相关性、主题建模、和其他现有技术。包含 Pro-ISIS 和 Anti-ISIS 推文的数据集的实验结果表明，我们的方法非常有希望在线自动识别恐怖分子内容。结果表明，与典型方法相比，使用诸如邻域重叠之类的图邻近度量来进行术语增强可以获得更高的精度、召回率和 F 度量。此外，我们发现应用基于时间的分析和术语扩充来识别激进账户提高了调查过程的精确度。结果表明，与典型方法相比，使用诸如邻域重叠之类的图邻近度量来进行术语增强可以获得更高的精度、召回率和 F 度量。此外，我们发现应用基于时间的分析和术语扩充来识别激进账户提高了调查过程的精确度。结果表明，与典型方法相比，使用诸如邻域重叠之类的图邻近度量来进行术语增强可以获得更高的精度、召回率和 F 度量。此外，我们发现应用基于时间的分析和术语扩充来识别激进账户提高了调查过程的精确度。