Reduced-Bias Co-trained Ensembles for Weakly Supervised Cyberbullying Detection

Abstract

Social media reflects many aspects of society, including social biases against individuals based on sensitive characteristics such as gender, race, religion, physical ability, and sexual orientation. Machine learning algorithms trained on social media data may therefore perpetuate or amplify discriminatory attitudes against various demographic groups, causing unfair decision-making. One important application for machine learning is the automatic detection of cyberbullying. Biases in this context could take the form of bullying detectors that make false detections more frequently on messages by or about certain identity groups. In this paper, we present an approach for training bullying detectors from weak supervision while reducing the degree to which learned models reflect or amplify discriminatory biases in the data. Our goal is to decrease the sensitivity of models to language describing particular social groups. An ideal, fair language-based detector should treat language describing subpopulations of particular social groups equitably. Building on a previously proposed weakly supervised learning algorithm, we penalize the model when discrimination is observed. By penalizing unfairness, we encourage the learning algorithm to avoid unfair behavior in its predictions and achieve equitable treatment for protected subpopulations. We introduce two unfairness penalty terms: one aimed at removal fairness and another at substitutional fairness. We quantitatively and qualitatively evaluate the resulting models’ fairness on a synthetic benchmark and data from Twitter comparing against crowdsourced annotation.