A Novel Weighted Soft Voting Method for Fake News Detectionin Social Media

Abstract

Social media platforms have brought about a massive change in the way information is created, shared and used. This accessibility, which is widespread, has also accelerated the dissemination of fake news and created public confusion and the presence of misinformation that has consequences on public life. Lexical features are the main ones used in conventional machine learning methods and are limited in their ability to detect features of the news content because they do not have the ability to identify the complex semantic and contextual properties. In this regard, this study suggests a Hybrid Ensemble Learning (ML) Algorithm for Fake News identification Using Semantic and Contextual Features to solve some of these issues. The suggested model combines complete Text Processing with feature engineering to capture lexical, semantic, syntactic, sentiment and contextual information from news articles. Sentence-BERT embeddings are used to create a semantic representation of each sentence, and contextual features, such as sentiment polarity, readability indices, linguistic patterns, and metadata-based attributes, are also integrated to improve the discriminative ability. Redundant information and increasing computational efficiency are achieved by using Recursive Feature Elimination (RFE) with feature selection by Mutual Information. A hybrid ensemble model of Extreme Random Forest (RF), Light Gradient Boosting Machine (LightGBM), Gradient Boosting (XGBoost), CatBoost and Random Forest (RF) classifiers with weighted soft voting strategy algorithm is used to classify selected features. Implementation of the framework is tested with benchmark fake news datasets (FakeNewsNet) and performance compared with some traditional machine learning models. The experimental results presented show that the proposed hybrid ensemble framework provides better classification accuracy (over 98%), and enhances recall, precision, F1-score and ROC-AUC. Ensemble learning with semantic embeddings and contextual feature analysis improve the robustness, generalization ability and explainability of fake news detection. The proposed method is a viable, scalable and dependable one for combating misinformation in modern social media and will be useful in detecting fake news in real-time fashion by automated methods.