Seoul's Machine-Learning Pipeline for Screening Participatory Budgeting Proposals

Seoul has run one of the world's largest participatory budgeting (PB) programmes since May 2012, allocating roughly 50 billion Korean won (about 2.2% of the city budget) through a General Committee of 250 citizens, officials and civil-society representatives; in 2014 alone the process received 1,533 proposals, of which 352 were ultimately selected for funding.

To help handle this volume, researcher Bokyong Shin (University of Helsinki) published a 2026 study aiming to cut citizens' 'learning costs' through automated categorisation and summarisation, and 'compliance costs' through predictive feedback before formal deliberation.

The study analysed 21,833 de-duplicated proposals (from 25,415 originally submitted) spanning 2013–2021, combining unsupervised topic modelling using Sentence-BERT embeddings — identifying 17 major topics across two clusters (nine infrastructure-focused, eight social/welfare-focused) — with a supervised Random Forest classifier that predicted proposal selection.

The Random Forest classifier predicted proposal selection with an AUROC of 0.73, outperforming logistic regression, gradient boosting, SVM and decision-tree alternatives.

This remains a research-stage capability built on historical open data rather than a tool confirmed to be running inside Seoul's live PB workflow, so its screening role should be read as a validated proof of concept, not an operational system. The programme's underlying scale and governance structure are independently corroborated by Participedia's case documentation.