ABSTRACTThis study addresses the issue of sampling biases in social media data-driven communication research. The authors demonstrate how supervised machine learning could reduce Twitter sampling bias induced from “proxy-population mismatch”. Particularly, this study used the Random Forest (RF) classifier to disentangle tweet samples representative of general publics’ activities from non-general—or institutional—activities. By applying RF classifier models to Twitter data sets relevant to four news events and a randomly pooled dataset, the study finds systematic differences between general user samples and institutional user samples in their messaging patterns. This article calls for disentangling Twitter user samples when ordinary user behaviors are the focus of research. It also builds on the development of machine learning modeling in the context of communication research.