Robust Cyberbullying Detection with Causal Interpretation

Cyberbullying poses serious threats to preteens and teenagers, therefore, understanding the incentives behind cyberbullying is critical to prevent its happening and mitigate the impact. Most existing work towards cyberbullying detection has focused on the accuracy, and overlooked causes of the outcome. Discovering the causes of cyberbullying from observational data is challenging due to the existence of confounders, variables that can lead to spurious causal relationships between covariates and the outcome. This work studies the problem of robust cyberbullying detection with causal interpretation and proposes a principled framework to identify and block the influence of the plausible confounders, i.e., p-confounders. The de-confounded model is causally interpretable and is more robust to the changes in data distribution. We test our approach using the state-of-the-art evaluation method, causal transportability. The experimental results corroborate the effectiveness of our proposed algorithm. The purpose of this study is to provide a computational means to understanding cyberbullying behavior from observational data. This improves our ability to predict and to facilitate effective strategies or policies to proactively mitigate the impact of cyberbullying.