Hybrid transformer-based recommender system for political news

Abstract

This study investigates a nudging-based news recommender designed to broaden exposure to political articles while preserving user satisfaction. We built a hybrid transformer-based recommender system that combines content-based embeddings (via a custom transformer architecture) with click-behavior signals to refine user profiles. A dedicated profile extension module augments each profile with semantically related concepts, subtly steering recommendations toward political news according to user-existing interests. In an online experiment with 168 participants, our nudging system significantly outperformed a popularity-based baseline (satisfaction: +12.84%, political clicks: +15.35%) and an interest-based baseline (satisfaction: +6.96%, political clicks: +6.44%). Notably, participants with the lowest initial political interest exhibited the largest engagement gains (clicks: +64.54% over popularity, +22.75% over interest) without compromising user satisfaction.