BITE-Net: A Multi-Feature Deep Learning Framework for Robust Clickbait Detection Integrating Psycholinguistic Signals, Semantic Contrast Modeling, and Multi-Head Self-Attention Fusion

Abstract

Contemporary digital platforms are experiencing a rise in the threat posed by purposefully crafted clickbait headlines, which take advantage of psychological vulnerabilities such as curiosity gaps, emotional triggers, and urgency cues, thereby compromising information integrity and user trust. The existing methods of detection use surface-based lexical patterns or independent transformer structures, which do not involve the underlying psychological processes of manipulations and propagation dynamics of the behavioral patterns that contain the deceptive content. In this paper, BITE-Net (Bi-directional and Integrated Trait Ensemble Network) refers to an advanced multi-feature deep learning architecture that incorporates NRC Emotion Lexicon psycholinguistic vectors, Semantic Contrast Vectors (SCV), Hyperbolic Weighting Scores (HWS), and Structural Virality Proxy (SVP) into a hybrid CNN-BiGRU-DeBERTa-v3. It involves a heterogeneous feature fusion system to dynamically combine heterogeneous feature representations, which are then computed using attention-based dimensionality reduction and finally classified into binary classification. BITE-Net-HO (headline-only, input-fair) achieves 97.10% on Kaggle 32K and 96.90% on Webis 2017, which is competitive with or exceeding all headline-only baselines, but full BITE-Net with article-level SCV has a 98.50% and 98.22% upper-bound reported separately. The contribution of psycholinguistic features is quantified through an interpretability analysis based on SHAP, and the superiority is tested with strict validity through McNemar statistical significance analysis. Detailed ablation experiments were conducted to ensure that each architectural element ensures the performance of robust detection.