Long-Term Trend Analysis for Social Media Content Published During COVID-19 Pandemic

ABSTRACT

Social media platforms are open for society to distribute numerous vectors of information upon personal interest. A substantive component of social media platforms is mining long-term trends. COVID-19 is an indispensable keyword captivating from last year in multiplex social media platforms providing opportunities for societal to express opinions in multiple languages. Existing literature is scarce in multiple aspects, such as the lack of resources that provide high accuracy long-term trend detection in social media platforms and missing the substantive factor for long-term trend detection, and time series forecasting over Natural Language Processing. This research presents a novel methodology to extract insights into low resource language and the responses are validated using the built crowdsourcing platform. Additionally, research employees with several classical time series analysis techniques and different LSTM RNN to identify the best-performing long-term trend analysis method. Novel concept deployed by extracting public data available in YouTube and Twitter which has at least one year of life expressing content related to COVID-19. Preponderance textual engagements were felt under the low resource language category. Crowdsourcing supports mitigating the risk of the intimidating issue. Finally, the data set is employed with time series forecasting. Mean Square Error and Huber loss were employed for testing. Authors argue there is no long-term trend in Twitter posts and YouTube videos with comments more petite than a hundred. YouTube videos with more than a hundred comments were further analyzed and the authors revealed that Stacked LSTM outperforms the LSTM used for trend analysis for more than one year by achieving 84.96% accuracy. © 2022 IEEE.