A HYBRID DEEP LEARNING BASED ENSEMBLE MODEL FOR EFFICIENT EARLY FAKE NEWS IDENTIFICATION: ENHANCING INFORMATION CREDIBILITY IN SOCIAL MEDIA
Alok Mishra, Halima Sadia
Department of Computer Science and Engineering
Integral University, Lucknow, Uttar Pradesh, India
Abstract: Ability to detect fake news with precision and speed is crucial in this age of rapid information dissemination online. Transformer- or sequence-based models don’t always function with full-text information, so you may not be able to utilize them initially or in real time. To overcome obstacles, study demonstrates how to implement a Contextual-Sequential-Ensemble Hybrid (CSE-Hybrid) approach. It uses XGBoost ensemble learning, an attention-gated fusion mechanism, BiLSTM-driven sequential dependency modeling, and BERT-based contextual encoding. The suggested technique includes a new Hybrid Early Detection Mechanism (HEDM) that employs multi-prefix sampling and confidence-based inference to allow for categorization of text inputs that are streaming or broken up. We employed three well-known datasets in this study: AG News, LIAR, and FakeNewsNet. We used a strict 5-fold nested cross-validation method that included bootstrap confidence intervals and statistical significance analysis. The CSE-Hybrid model had problems with blank text, but it did better than baseline approaches on other measures including F1-score and AUC. It all made sense when I used attention heatmaps, t-SNE visualizations, and error location analysis. The time-to-detection statistic is one way to tell how well the model works. It seems to be almost flawless with just 65–75% of the text. We also spoke about ethical deployment tactics, making fair datasets, and other relevant concerns to keep everyone safe and up to date. The CSE-Hybrid model provides a reliable, user-friendly, and context-aware framework for quickly and accurately finding bogus news. The proposed approach employs attention-gated fusion, early-detection learning, and confidence-aware inference, distinguishing itself from rival frameworks that rely on component selection.
Keywords: BERT, BiLSTM, XGBoost, Text Classification, Fake News Detection, Early Detection, Hybrid Deep Learning.
VOLUME 10 ISSUE 03 2026: 94 – 117