ISSN :2582-9793
  • Subscribe

    • Breaking the Bias: Deep Learning Meets Hybrid Sampling for Cyberbullying Detection

    Original Research (Published On: 17-Dec-2025 )
    DOI : https://doi.org/10.54364/AAIML.2025.54262

    Mavara Malik, Malik Muhammad Missen and Hijab Asad

    Adv. Artif. Intell. Mach. Learn., 5 (4):4732-4746

    1. Mavara Malik: The Islamia University Bahawalpur, Bahawalpur.

    2. Malik Muhammad Missen: The Islamia University Bahawalpur, Bahawalpur.

    3. Hijab Asad: The Islamia University Bahawalpur, Bahawalpur.

    Download PDF Here

    DOI: 10.54364/AAIML.2025.54262

    Article History: Received on: 30-Sep-25, Accepted on: 10-Dec-25, Published on: 17-Dec-25

    Corresponding Author: Mavara Malik

    Email: mavaramalik1@gmail.com

    Citation: Mavara Malik, et al. Breaking the Bias: Deep Learning Meets Hybrid Sampling for Cyberbullying Detection. Advances in Artificial Intelligence and Machine Learning. 2025;5(4):262. https://dx.doi.org/10.54364/AAIML.2025.54262


    Abstract

        

    Cyberbullying detection faces dual challenges: severe class imbalance (minority <15%)
    and systematic demographic bias. These problems are deeply intertwined—models trained
    on imbalanced data exploit spurious correlations between demographic markers and labels.
    This work presents an integrated framework addressing both jointly. Bias-aware hybrid
    sampling combines SMOTE-Tomek and ADASYN with demographic tracking, reducing
    imbalance from 6.79:1 to 1.28:1. A dual CNN-BiLSTM with 8-head attention operates
    over 492d enriched embeddings. Composite loss with adaptive weighting incorporates
    focal loss, demographic parity penalties, and equalized odds constraints.
    Evaluation on 448,873 samples shows 88.9% F1 with 86.7% minority recall. Demographic
    parity improved 96.7% (from −0.212 to −0.0055), equalized odds 77.1%, disparate
    impact 41.8%. Cross-platform validation shows F1 degradation ≤ 0.6%. Versus
    transformers: 9× fewer parameters (12M vs. 110M+), 8× faster inference (23ms vs.
    187ms), 94% superior fairness (p < 0.0003). Results challenge conventional fairnessaccuracy
    tradeoffs.

    Statistics

       Article View: 939
       PDF Downloaded: 8