Population-Scale Network Embeddings Expose Educational Divides in Network Structure Related to Right-Wing Populist Voting

A team of researchers has published a groundbreaking study using machine learning to analyze population-scale social networks and their connection to political behavior. The team, led by Malte Lüken and including Javier Garcia-Bernardo, Sreeparna Deb, Flavio Hafner, and Megha Khosla, created network embeddings for all 17 million+ people in the Netherlands using administrative registry data. They constructed a massive network representing five key shared social contexts: neighborhood, work, family, household, and school, then used machine learning techniques to encode these connections into numerical representations called embeddings that capture each individual's position within the broader social structure.

The researchers then tested whether these network embeddings could predict right-wing populist voting behavior. While the embeddings alone predicted voting above chance level, they performed worse than traditional individual characteristics like demographics. However, when the researchers transformed the embeddings to make their dimensions more sparse and orthogonal—essentially making them more interpretable—they discovered one particular embedding dimension that showed strong association with voting outcomes. This dimension, when mapped back to the population network, revealed that differences in educational ties and attainment corresponded to distinct network structures that were associated with right-wing populist voting patterns.

The study represents a significant methodological advancement in making population-scale network embeddings interpretable, moving beyond black-box predictions to uncover meaningful social patterns. Substantively, it provides evidence linking structural differences in educational networks to political behavior at a national scale, offering new insights into how social connections and educational segregation might influence voting patterns in modern democracies.