Candia Vallejo, CristianFuentes Jofré, Adolfo Ignacio2026-08-032026-08-032026https://hdl.handle.net/11447/10948A dissertation submitted to the Faculty of Government at the Universidad del Desarrollo in partial fulfillment of the requirements for the degree of Doctor in Social Complexity SciencePolarization is often described as a defining feature of contemporary politics, yet most evidence comes from political elites or from studies focused on the United States. The structure of citizens’ ideological preferences, and how it varies across societies, remains poorly understood. We present DYNAMAP, a transparent neural framework that infers the latent geometry of human value organization directly from behavioural data. By embedding individuals and policy options in a shared space based on pairwise comparisons among real public policy issues, the model captures how citizens balance competing political values without relying on self-reports or elite cues. Applied to revealed-preference datasets from Chile, France, and Brazil, three contrasting political contexts, DYNAMAP identifies a shared, low-dimensional ideological structure anchored by a left–right axis and a secondary component reflecting contextual diversity. Within this common geometry, polarization follows distinct trajectories: among younger generations it intensifies in France and Brazil but weakens in Chile, driven by moderation among younger right-leaning participants. Traditional approaches based on self-declared ideology provide limited insight into how citizens actually organize competing values. Given that DYNAMAP learns ideological structure directly from behavioural comparisons, it reveals context-specific generational patterns that remain obscured in conventional survey data. These findings suggest that polarization reflects context-dependent rearrangements within a stable ideological space, offering a scalable and interpretable approach for understanding how societies organize disagreement.90 p.enPolarización políticaPreferencias reveladasIdeologíaAprendizaje automáticoComparaciones pareadasPolitical polarizationRevealed preferencesIdeologyMachine learningPairwise comparisonsThesis