LATIN HYPERCUBE SAMPLING HIPPOPOTAMUS OPTIMIZATION (LHSHO) BASED FEATURE SELECTION AND STACKING ENSEMBLE DEEP LEARNING (SEDL) MODEL FOR FAKE NEWS DETECTION

Abstract

Fake news's online development through platforms has become an important concern, affecting political stability, public opinion, and the spread of trustworthy information. Machine learning (ML) and deep learning (DL) techniques are introduced to identify and classify fake news. Several features are presented in the detection model, features are manually extracted from news, and the input text includes lengthy paragraphs. It becomes reduces the detection accuracy of the model. To solve this issue, in this paper novel summarization, and fake news detection method is introduced for social networks. Firstly, a fake news dataset is gathered from BuzzFeed and PolitiFact. Secondly, Hybrid Bidirectional Encoder Representation from Transformers (HBERT) is developed for text summarization to consider both fine-grained text information and label semantics at an identical period. Thirdly, text preprocessing and feature extraction are performed for summarized text. Fourthly, Latin Hypercube Sampling Hippopotamus Optimization (LHSHO) is introduced for feature selection which includes the hippopotamus natural reaction of running away from predators and dynamically attempting to distance itself from reduced error. LHS will help to prevent the initial population generation, and solves local optimal problem during selection process. Lastly, the Stacking Ensemble Deep Learning (SEDL) model is introduced, which combines various models such as Conditional Generative Adversarial Network (CGAN), Self Adaptive Attention-based bidirectional Long Short-Term Memory (SAA-BiLSTM), and Stacked Sparse Auto Encoder (SSAE). Projection-based extension of GAN is called CGAN. SSAE, a sparse penalty term is employed to achieve more accurate and effective detection. Stacking is used to combine the results of these classifiers by training a meta-model (Logistic Regression (LR)). Performance is assessed using metrics including precision, recall, F-measure, and accuracy.