Estimating Emotion Contagion on Social Media via Localized Diffusion in Dynamic Graphs
We present a computational approach for estimating emotion contagion on social media networks. Built on a foundation of psychology literature, our approach estimates the degree to which the perceivers' emotional states (positive or negative) start to match those of the expressors, based on the latter's content. We use a combination of deep learning and social network analysis to model emotion contagion as a diffusion process in dynamic social network graphs, taking into consideration key aspects like causality, homophily, and interference. We evaluate our approach on user behavior data obtained from a popular social media platform for sharing short videos. We analyze the behavior of 48 users over a span of 8 weeks (over 200k audio-visual short posts analyzed) and estimate how contagious the users with whom they engage with are on social media. As per the theory of diffusion, we account for the videos a user watches during this time (inflow) and the daily engagements; liking, sharing, downloading or creating new videos (outflow) to estimate contagion. To validate our approach and analysis, we obtain human feedback on these 48 social media platform users with an online study by collecting responses of about 150 participants. We report users who interact with more number of creators on the platform are 12% less prone to contagion, and those who consume more content of `negative' sentiment are 23% more prone to contagion. We will publicly release our code upon acceptance.