Socio-spatial Self-organizing Maps: Using Social Media to Assess Relevant Geographies for Exposure to Social Processes

Social Media offers a unique window into measuring attitudes like racism and homophobia, which could be relevant social determinants in human immunodeficiency virus (HIV) risk. However, individual Tweets can be noisy and existing areas by which exposures are measured, like Zip codes, average over administratively-defined boundaries, limiting the use of the precise geo-information linked to social media. Hence, we need a method to identify relevant, homogeneous areas of social attitudes. To address this, we augment traditional selforganizing maps (SOMs), to topologically constrain the clusters, and return a controlled number of non-overlapping clusters. Our approach (called Socio-spatial-SOMs, “SS-SOMs”) uses neural embedding for text-classification and neural networks for clustering, to best identify regions of consistent social attitudes, semantically and geographically. We find that SS-SOMs generate homogeneous, well-defined and more topically-similar areas in comparison to traditional SOMs and Zip codes, and are robust to missing data. We demonstrate the impact of this new way to spatially represent social attitudes using mobility data from a cohort of men at high risk for HIV, finding that their exposure to racism and homophobia as measured using SS-SOMs differs by up to 42% as compared to using Zip code based measures.