AdaGraph-HATE for real-time gendered hate speech detection in multilingual social media

Abstract

Social networking sites process millions of messages daily, yet detecting implicit misogyny remains challenging due to sarcasm, cultural variation, and language-dependent expressions. Existing approaches demonstrate promising results but suffer from limited language adaptation and impracticality for real-time deployment. This paper introduces AdaGraph-HATE, a hybrid ensemble framework combining seven diverse classifiers SVM, KNN, logistic regression, ANN, CNN, BERT, and SBERT with a novel graph-aware dynamic weighting mechanism that adapts model weights based on confidence scores, inter-model diversity, and correlation structure captured through graph representations. Comprehensive evaluation across five datasets spanning eight languages with 124,456 labeled tweets reveals that AdaGraph-HATE achieves 90.02% accuracy and 0.7354 Macro F1, showing a small statistically significant improvement over SBERT (88.92% accuracy, 0.6753 Macro F1) with p < 0.001 but overlapping confidence intervals. However, critical analysis exposes substantial trade-offs: hate class detection F1 drops from 0.2867 (SBERT) to 0.1102 (AdaGraph) a 61.6% relative decrease while severe misclassifications increase by 82.6%. Cross-dataset evaluation further reveals limited generalization, with AdaGraph-HATE underperforming SBERT on four of five datasets. From a systems perspective, the framework demonstrates production-ready scalability, processing 18,234 tweets per second with 68.4 ms P95 latency and costing $0.19 per million tweets. These findings indicate that while adaptive multi-model integration offers throughput and cost benefits, it introduces significant minority-class detection trade-offs. SBERT remains superior for safety–critical moderation, while AdaGraph-HATE suits high-volume environments where training and deployment distributions match. This study provides both theoretical insights into ensemble design and practical guidance for deploying reliable hate speech detection systems in real-world scenarios.