Half-time seminar by Matias Piqueras: "Analysis of Online Visual Political Communication"

Abstract:

The proliferation of online platforms has generated a massive amount digital traces of visual political communication. Combining computer vision and Bayesian statistics, this presentation introduces novel and scalable computational methods to measure theoretically relevant social science constructs from large-scale multimodal data.

First, we address the categorisation of visual content by introducing the Visual Structural Topic Model (vSTM), a method integrating pretrained image embeddings with a structural topic model. This approach captures the semantic complexity of images and allows for the analysis of how visual topics relate to covariates. We demonstrate its application in a study of the 2019 Global Climate Strike on Twitter/X, where we identify polarisation and analyse how visuals are contested by movement and countermovement actors.

Second, we develop a model for inferring the ideology of social media posts and accounts from a graph of behavioural signals. The Poisson Factorisation Point Ideal (PFIP) model is a scalable Bayesian framework that extends the canonical ideal point model by combining a kernel approximation to ideological distance, with a factorisation of the log likelihood. This makes fitting the model to graphs of billions of interactions tractable, as the time and memory complexity is proportional to the number of non-zero elements in the adjecency matrix, which is typically very sparse.