Platform-aware misogyny detection on social media: a comparative analysis of class imbalance, text normalisation, and model selection

Online misogyny remains a major problem across social media, yet most detection systems are developed and evaluated on a single platform, raising questions about whether successful design choices generalise across platforms. This study addresses that gap through a multi-platform analysis of misogyny detection on X (Twitter), Reddit, and a combined dataset, covering three imbalance ratios (IR = 1.07, 8.38, and 3.33). Results show that design choices are highly platform dependent and can have opposite effects. Class balancing provides minimal benefit on the near-balanced X (Twitter) dataset but improves misogyny F1 by over 18% on the highly imbalanced Reddit dataset. In contrast, synonym-based augmentation consistently reduces performance, suggesting lexical substitutions distort context-sensitive misogynistic cues across platforms. Model rankings also vary by platform. BERT performs best on individual homogeneous platforms (85% F1 on Reddit; 78.6% on X (Twitter)), while HateBERT proves more robust on the heterogeneous combined dataset (82.6–83.8% F1). Transformer models generalise better than FastText to unseen misogynistic expressions, but all models struggle with implicit, humor-driven, and stereotype-based misogyny, highlighting implicit detection as an ongoing challenge.