Who Falls into the Echo Chamber? Identifying Risk Features of Social Media Users via Explainable Machine Learning

Echo chambers are increasingly recognized as a critical challenge in online information environments where algorithmic filtering and user preferences limit exposure to diverse viewpoints. While existing studies have primarily relied on content- or network-based methods to detect echo chambers, few have focused on identifying echo chambers using machine learning methods and few have looked at the user risk features that predict participation in an echo-chamber. This study proposes a machine learning framework to predict echo chamber membership based on user features drawn from two representative platforms, Weibo and Twitter (X). We utilized a published dataset covering discussions on the Omicron variant and the Tokyo 2020 Olympics, incorporating 20 user features, including 15 general behavioral features and 5 Big Five personality traits. Classification models were developed using three advanced algorithms: Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost). Feature contributions were interpreted using Shapley Additive Explanations (SHAP) and CatBoost was the more efficient (or effective or fast). Results revealed that strong emotional polarity, frequent social interaction, and high neuroticism were key user risk features associated with echo chamber participation across both platforms. The results matter because they move beyond traditional content or network-based explanations of echo chambers, shedding light on the individual level psychological and behavioral traits that make users more vulnerable to these environments. This user-centric perspective allows for a deeper understanding of who is at risk and why, which is critical for designing interventions that don’t just target what people see, but how and why they engage with information in the first place.