We proposed an unsupervised method uses Author-Topic Modeling to extract topics of each states' tweets, then make predictions based on states' topic distributions. Our methods achieved accuracy approaching human expert's and outperformed supervised classification of FastText.
Reference:
[1]Eisenstein, Jacob, et al. "A latent variable model for geographic lexical variation." Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2010.