Rumor identification and diffusion impact analysis in real-time text stream using deep learning

Researchers have applied marvelous endeavors in developing and designing methodologies to understand the behavioral of social media data. Various social media data have allegedly been used as a platform of rumor promotions with momentous political, economic, personal and social outcomes in numerous countries. In the existing approaches, diffusion rate of rumor words in terms of continuous and discrete time is not analyzed on the basis of their contextual polarity including their impact on social media users. It has been observed that the trust on rumor words based on user classification also plays a significant role which needs to be addressed. Therefore, we designed rumor detection and analysis methodologies in continuous and discrete time via deep learning. An efficient preprocessing approach is designed to clean the collected data. A weigh-based threshold mechanism is opted to compute various rumor-based features and their diffusion tendency in the continuous and discrete case. The proposed approach is tested on five sets of real-time data and on comparison with state-of-the-art (SOTA) methods which are CNN, KNN, LSTM and Word2Vec and noticed that rumor keyword recall and rumor keyword precision has been significantly improved. Efficiency of the proposed approach in terms of rumor-based keyword detection is also computed in which proposed approach detects 84.58, 83.46, 83.59, 83.05 and 84.47 with less loss of data than existing SOTA methods. Computation time of proposed approach is 0.59, 0.53, 0.57, 0.49, 0.47 for CoAID, FakeNewsNet, LIAR, CORD-19 and PHEME, respectively, which is quite less than SOTA methods (LSTM, CNN, Word2Vec and KNN). Practically, our proposed approach will play a significant role to improve the eminence of the rumor detection mechanism for social media platforms. Further, if we categorize the people on the basis of their influential role to spread rumor, then the trust factor based on user’s classification is computed. A real-time case study is an evidence of the efficiency of our proposed approach which provide empirical findings of detecting online rumor including trust based on social media user crowd. The source code is available at the GitHub.