Variance of Twitter Embeddings and Temporal Trends of COVID-19 cases

TL;DR

This paper explores how Twitter data and word embeddings can be used to predict COVID-19 case surges with a lead time of up to 30 days, aiding timely resource planning.

Contribution

It introduces a novel method leveraging Twitter embeddings and Significant Dimensions to forecast COVID-19 case increases with high accuracy.

Findings

Predicts COVID-19 case surges with 15 and 30-day lead times

Achieves R2 scores of 0.80 and 0.62 for predictions

Identifies thematic significance of embedding dimensions

Abstract

The severity of the coronavirus pandemic necessitates the need of effective administrative decisions. Over 4 lakh people in India succumbed to COVID-19, with over 3 crore confirmed cases, and still counting. The threat of a plausible third wave continues to haunt millions. In this ever changing dynamic of the virus, predictive modeling methods can serve as an integral tool. The pandemic has further triggered an unprecedented usage of social media. This paper aims to propose a method for harnessing social media, specifically Twitter, to predict the upcoming scenarios related to COVID-19 cases. In this study, we seek to understand how the surges in COVID-19 related tweets can indicate rise in the cases. This prospective analysis can be utilised to aid administrators about timely resource allocation to lessen the severity of the damage. Using word embeddings to capture the semantic meaning of tweets, we identify Significant Dimensions (SDs).Our methodology predicts the rise in cases with a lead time of 15 days and 30 days with R2 scores of 0.80 and 0.62 respectively. Finally, we explain the thematic utility of the SDs.