Cyberbullying detection using probabilistic socio-textual information fusion

Cyberbullying is an important socio-technical challenge in Online Social Networks (OSN). With the growth trends of heterogeneous data in OSN, better network characterization, and textual feature sophistication, recent efforts have realized the value of looking at heterogeneous modes of information including textual features, social features, and image-based features for better cyberbullying detection. These approaches, however, still use these features either individually or combine them `naively' without considering the different confidence levels associated with each feature or the interdependencies between features. We propose a novel probabilistic information fusion framework that utilizes confidence score and interdependencies associated with different social and textual features and uses those to build better predictors for cyberbullying. The performance of the proposed approach was compared to a recent approach in literature which used a similar dataset and features and the proposed approach resulted in significant improvements in terms of cyberbullying detection.