Fake News Identification on Social Platforms Using Extreme Gradient Boosting in Information Filtering Systems

The proliferation of algorithmically amplified misinformation across social media platforms has produced measurable harms at societal scale: the Reuters Institute Digital News Report 2024 documents that 56% of adults in 46 countries express concern about distinguishing real from fabricated online news, while the World Health Organization estimates that the 2020-2024 COVID-19 infodemic-false information spreading concurrently with the pandemic-contributed to vaccine hesitancy affecting an estimated 318 million people globally. Automated fake news identification systems capable of operating at platform speed and scale are consequently a research priority of high practical urgency. This paper evaluates Extreme Gradient Boosting (XGBoost) for automated fake news identification, applied to a multi-platform, multi-modal corpus of 248,400 social media news items drawn from four platforms (Twitter/X: 84,600; Facebook: 72,400; Reddit: 48,200; Telegram: 43,200) over a 36-month collection window (January 2022-December 2024). Fifty-eight features across three modalities are engineered: textual and linguistic features (24 features), social graph and propagation features (18 features), and user behaviour and account features (16 features). XGBoost is configured with 400 estimators, maximum tree depth 7, learning rate 0.08, subsample ratio 0.84, column subsampling ratio 0.72, and early stopping at 42 rounds, achieving$\mathbf{F 1}=\mathbf{0. 8 8 4}$and$\mathbf{A U C}-\mathbf{R O C}=\mathbf{0. 9 4 2}$on the stratified$\mathbf{2 0 \%}$held-out test set$(\mathbf{n}=\mathbf{4 9, 6 8 0})$-outperforming Logistic Regression (F1: 0.724), SVM (F1: 0.784), Random Forest (F1: 0.801), and BERT-base binary classifier (F1: 0.848). SHAP attribution identifies the linguistic credibility score (22.4% of SHAP variance), source reputation index (18.8%), propagation velocity (16.4%), user account age (12.8%), and sentiment divergence (10.4%) as the five primary detection features, collectively accounting for 80.8% of model output variance. Deployed as a real-time content filtering layer across the four platforms, the XGBoost system processes each news item in 0.024 seconds, enabling preamplification interception at platform speed, and reduces the downstream reach of identified fake news items by 64.8% relative to the unfiltered baseline over the evaluation window.

Hali tarjima qilinmagan