Towards Fair and Accurate Short VAAs: Adaptive Ordering and Imputation in Democratic Recommendation Systems

Abstract

In increasingly fragmented party systems, voters face growing difficulties in making informed electoral decisions. Voting Advice Applications (VAAs), questionnaires consisting of political statements, have become central democratic tools by helping citizens compare their political preferences with those of competing parties. Yet, modern VAAs face a fundamental trade-off: while shorter questionnaires could reduce user fatigue and dropout, they also suffer from severe advice degradation because they consider fewer statements. Existing solutions using adaptive questionnaires and imputation show promise, but remain largely focused on candidate-based systems, which only make up a small fraction of the overall electoral systems, and rarely address fairness across parties and opinion groups. This paper introduces a novel adaptive framework for party-based VAAs that combines active learning with imputation trained solely on previous user responses. Rather than relying on a fixed questionnaire, our system dynamically personalizes which statements users receive, in what order, and how many are needed before reliable voting advice can be generated. Evaluating the approach on two German municipal VAA datasets from Saarbrücken and Halle (Saale), we show that imputation and adaptive item selection strategies improve recommendation quality by more than 12 percentage points compared to conventional short VAAs. Beyond performance gains, our findings reveal systematic inequalities in how short VAAs represent minority opinions and smaller parties, highlighting the importance of cross-party fairness in the design of democratic digital tools.