The memory cycle of time-series public opinion data: Validation based on deep learning prediction

Abstract

This study explored the response mechanism of time-series public opinion data to sustained external stimuli based on COVID-19 pandemic and online news data (N = 39,753). We employed multiple deep learning models (CNN1D,CNN2D,LSTM,CNN1D-LSTM, and CNN2D-LSTM), using pandemic spread data over varying time windows (window ∈ [5–100]) as input to predict public opinion trends at different lag periods (lag ∈ [1–100]). Subsequently, we investigated the response mechanism of time-series public opinion data to external stimuli by observing the distribution patterns of effective combinations of input windows and lag periods. The findings indicate that statistical data on COVID-19 transmission contain latent knowledge about future public opinion trends. In effective prediction patterns, the size of the data input window and effective lag periods exhibit stable periodicities. This periodicity presents different timescales in the short-, medium-, and long-term, highlighting the public's near-term feedback, ongoing concern, and long-term review regarding major social issues.