Enhancing Rumor Detection from Social Media Texts with an Ensemble Model for Residual Graph Convolutional Networks

Detecting rumors is the most difficult task when working with social media sites. To address this problem, various Deep Learning (DL)-based rumor detection methods have evolved in previous decades. Amongst, Bidirectional Encoder Representations from Transformers with Attention based Balanced Spatial-Temporal Residual Graph Convolutional Networks (BERT-ABSTRGCN) was developed to address the models convergence issues for rumor stance classification. However, the training procedure will slow down significantly due to complex features, which increase computational complexity and strain hardware resources needed for the large feature set. For this, Ensemble-based multiple time series analysis technique is developed for timely rumor recognition on social media by eliminating the feature set complexity on larger dataset. Initially, the collected tweet streams are pre-processed to obtain the clean data. Then, BERT is applied on the pre-processed data to extract different relevant tweet timestamps features without any time delay for models training. Reaction counts are used as features in time-series vectors that are created from Twitter conversations, which are lists of tweets. These vectors are then inputted into DL models to produce the time-series data. The suggested ensemble model improves the classification model performance-using majority voting across multiple ABSTRGCN within the ensemble, which leverages the individual strengths of each network to enhance overall performance. The proposed model reduces resource and computational complexity in model training on large datasets for improvising the rumor detection efficiency in social media. The complete model is termed as BERT-Ensemble of ABSTRGCN (BERT-EABSTRGCN). Furthermore, the findings show that the suggested model hits 96.89% accuracy on PHEME datasets respectively compared to the classical models.