A Multilayered Preprocessing Approach for Recognition and Classification of Malicious Social Network Messages

Abstract

The primary methods of communication in the modern world are social networks, which are rife with harmful messages that can injure both psychologically and financially. Most websites do not offer services that automatically delete or send malicious communications back to the sender for correction, or notify the sender of inaccuracies in the content of the messages. The deployment of such systems could make use of techniques for identifying and categorizing harmful messages. This paper suggests a novel multilayered preprocessing approach for the recognition and classification of malicious social network messages to limit negative impact, resulting in fewer toxic messages, scams, and aggressive comments in social media messages and commenting areas. As a result, less technical knowledge would be required to investigate the effects of harmful messages. The dataset was created using the regional Lithuanian language with four classes: aggressive, insulting, toxic, and malicious. Three machine learning algorithms were examined, five use cases of a multilayered preprocessing approach were suggested, and experiments were conducted to identify and classify harmful messages in the Lithuanian language.