In this paper, the problem of accurately detecting cyberbullying instances in a timely manner is addressed. In stark contrast to most prior work that attempts to detect aggressive behavior by looking at individual messages, we consider cyberbullying as a repeated aggressive behavior towards an individual and jointly examine multiple messages exchanged between users. Moreover, we are interested in reaching an accurate decision fast. To this end, we propose a novel hierarchical approach that (i) first characterizes an individual message as aggressive or not by evaluating the optimum least number of informative features extracted from this message, and (ii) uses this new knowledge to decide if it should continue reviewing messages or conclude the process and raise a cyberbullying alert. It is shown that the proposed approach is guaranteed to review the least number of messages before reaching a decision, while the optimum decision rule is shown to minimize the average Bayes risk. Evaluation on real-world Instagram data demonstrates that the proposed method is able to accurately detect cyberbullying instances by reviewing up to 59% less messages than the state-of-the-art.
展开▼
