Detection of Cyber Grooming in Online Conversation

In this paper we will focus on the detection of sexual predators in online chat conversations. We use 3 different approaches (message-based, author-based and conversation-based) combined with 5 different classification algorithms and 2 different features sets. The best results were obtained using either the author-based approach with the Neural Network classifier on the TF-IDF feature set, or the conversation-based approach using the Ridge or the Naïve Bayes classifier on the TF-IDF feature set. In this paper, for the first time, we looked at how quick a predator could be detected, and found that in most cases 26-161 messages of a conversation were sufficient. This constitutes only a small fraction of the full conversations, showing that we can have an early detection system of sexual predators instead of knowing in retrospect that a child was the victim of a sexual predator.