Word Clustering Using Graphical Lasso-Guided PCA for Trend Analysis of COVID-19

This paper proposes a word clustering method using graphical lasso-guided principal component analysis (PCA) for trend analysis of coronavirus disease (COVID-19). We define changes in daily frequencies of words on Twitter as trends, and clustering denotes to find similar trends. There is a problem that trends based on indirect correlations degrade the clustering performance. To address this problem, we newly develop graphical lasso-guided PCA. Specifically, graphical lasso is able to obtain a partial correlation matrix (a graph that represents direct correlations between trends). By calculating loadings of PCA to the partial correlation matrix (authority scores calculated by a hyperlink-induced topic search algorithm), accurate clustering becomes feasible. We conducted experiments by collecting Japanese tweets about COVID-19 from March 1, 2020 to April 30, 2020. The results show that our graphical lasso-guided PCA can distinguish two clusters before and after a state of emergency, unlike comparative method using indirect correlations.