Alliance Dynamics and Vote Outcome Prediction in the Brazilian Chamber of Deputies

Abstract

Legislative roll-call records provide a high-resolution view of how representatives align on institutionally consequential decisions. In multiparty settings such as the Brazilian Chamber of Deputies, however, coalition structure can be blurred by weakly polarized voting sessions and noisy similarity edges. Predictive performance can also be overestimated when temporal leakage is not carefully avoided. This Master’s dissertation presents a computational framework integrating two complementary goals. First, yearly voting-similarity networks and community detection identify cohesive blocs interpretable as political alliances. A noise-reduction strategy combines polarization-aware proposition filtering and weak-edge pruning before applying Leiden, enabling clearer communities and longitudinal analysis from 2004 to 2023. Second, VOTE-RAP formulates proposition-level outcome prediction under strict chronological evaluation. Results show a 10.95% modularity improvement over a backbone baseline and accurate leakage-safe forecasting (AUROC 0.908; rejected-class F1 0.703).