Characterizing Dietary Choices, Nutrition, and Language in Food Deserts via Social Media

Social media has emerged as a promising source of data for public health. This paper examines how these platforms can provide empirical quantitative evidence for understanding dietary choices and nutritional challenges in “food deserts” -- Census tracts characterized by poor access to healthy and affordable food. We present a study of 3 million food related posts shared on Instagram, and observe that content from food deserts indicate consumption of food high in fat, cholesterol and sugar; a rate higher by 5-17% compared to non-food desert areas. Further, a topic model analysis reveals the ingestion language of food deserts to bear distinct attributes. Finally, we investigate to what extent Instagram ingestion language is able to infer whether a tract is a food desert. We find that a predictive model that uses ingestion topics, socio-economic and food deprivation status attributes yields high accuracy (>80%) and improves over baseline methods by 6-14%. We discuss the role of social media in helping address inequalities in food access and health.