Predicting the demographics of Twitter users with programmatic weak supervision

Predicting the demographics of Twitter users has become a problem with a large interest in computational social sciences. However, the limited amount of public datasets with ground truth labels and the tremendous costs of hand-labeling make this task particularly challenging. Recently, programmatic weak supervision has emerged as a new framework to train classifiers on noisy data with minimal human labeling effort. In this paper, demographic prediction is framed for the first time as a programmatic weak supervision problem. A new three-step methodology for gender, age category, and location prediction is provided, which outperforms traditional programmatic weak supervision and is competitive with the state-of-the-art deep learning model. The study is performed in Flanders, a small Dutch-speaking European region, characterized by a limited number of user profiles and tweets. An evaluation conducted on an independent hand-labeled test set shows that the proposed methodology can be generalized to unseen users within the geographic area of interest.