Content-driven cyber community identification using deep learning

Abstract Cyber communities often lack clear geographic or formal boundaries, which makes them difficult to characterize and compare using traditional community-detection assumptions. Most prior work either relies on network structure or reports content analytics without clearly stating how such outputs can be used to distinguish communities over time. Here we propose a content-centric framework that characterizes cyber communities implicitly through aggregate signals computed from user-generated content (UGC). Specifically, we train two LSTM-based models: an 11-class topic classifier (31,386 sample) and a 3-class sentiment classifier (162,980 sample). We then aggregate model outputs at the cohort level to derive community signatures based on topic-mixture distributions, sliding-window sentiment trajectories, and interpretable statistics including entropy-based content diversity and TF–IDF salience. The proposed topic and sentiment models achieve 94.08% and 93.8% accuracy, respectively, with strong precision, recall, and F1-scores, enabling reliable extraction of community-level profiles. These profiles support monitoring and comparison of cyber communities by highlighting dominant themes, shifts in sentiment over time, and changes in content diversity. This framework provides an interpretable and low-overhead approach for tracking community dynamics and can support downstream applications such as marketing analytics, harassment monitoring, and community management.