Measuring Gender: A Machine Learning Approach to Social Media Demographics and Author Profiling

Social media has become a preeminent medium of communication during the early 21(st) century, facilitating dialogue between the political sphere, businesses, scientific experts, and everyday people. Researchers in the social sciences are focusing their attention on social media as a central site of social discourse, but such approaches are hampered by the lack of demographic data that could help them connect phenomena originating in social media spaces to their larger social context. Computational social science methods which use machine learning and deep learning natural language processing (NLP) tools for the task of author profiling (AP) can serve as an essential complement to such research. One of the major demographic categories of interest concerning social media is the gender distribution of users. We propose an ensemble of multiple machine learning classifiers able to distinguish whether a user is anonymous with an F1 score of 90.24%, then predict the gender of the user based on their name, obtaining an F1 score of 89.22%. We apply the classification pipeline to a set of approximately 44,000,000 posts related to COVID-19 extracted from the social media platform Twitter, comparing our results to a benchmark classifier trained on the PAN18 Author Profiling dataset, showing the validity of the proposed approach. An n-gram analysis on the text of the tweets to further compare the two methods has been performed.