Abstract
Numerous studies have been concerned with developing new authorship recognition systems to address the increasing rates of cybercrimes associated with the anonymous nature of social media platforms, which still offer the opportunity for the users not to reveal their true identities. Nevertheless, it is still challenging to identify the real authors of social media's offensive and inappropriate content. These contents are usually very short; therefore, it is challenging for stylometric authorship systems to assign controversial texts to their real authors based on the salient and distinctive linguistic features and patterns within these contents. This research introduces a new stylometric authorship system that considers both the shortness of data and the peculiar linguistic properties of Arabic. A corpus of 20, 357 tweets from 134 Twitter users. A document clustering based on Document Index Graph (DIG) model was used to classify input patterns in the tweets that shared common linguistic features. A comparative analysis using Vector Space Clustering (VSC) model based on the Bag of Words (BOW) model, conventionally used in authorship recognition applications, was used. Results indicate that the proposed system is more accurate than other standard authorship systems mainly based on vector space clustering methods. It was also clear that the model had the advantage of providing complete information about the documents and the degree of overlap between every pair of documents, which was useful in determining the similarity between documents.