Demographic Inference from Social Media Data with Multimodal Foundation Models: Strategies, Evaluation, and Benchmarking

Abstract

Contemporary social media platforms enable users to generate and share vast amounts of behavioral data and opinions at an unprecedented scale and granularity. These social media data have been increasingly used in research across various domains, including public opinion analysis(Dong & Lian, 2021; Tavoschi et al., 2020), political communication(Liu et al., 2021), disaster response(Houston et al., 2015), health surveillance(Skaik & Inkpen, 2021; Yang et al., 2023), misinformation detection(Shu et al., 2017), and event monitoring(Mredula et al., 2022). However, using these data to understand social phenomena at the population level remains challenging due to their inherent non-representativeness and sampling bias. Social media users rarely disclose explicit demographic attributes such as age, gender, or race. Without this contextual information, analyses of online attitudes, opinions, and behaviors can be biased toward dominant or more vocal groups, obscuring systematic differences across social strata.