Best Intervention for Recommender Systems

Abstract

Social media companies deploy algorithms to deliver targeted content to their users, but these algorithms are often criticized for spreading fake news and increasing polarization. To address these issues, the literature attempts to minimize polarization-related metrics, but the complex nature of these tasks often requires problem-specific hand-tailored algorithms and heuristics. We introduce a novel gradient-based approach to solve a broad class of bilevel optimization problems, which includes a wide range of Friedkin-Johnsen opinion dynamics problems. Our algorithm demonstrates inherent flexibility and scalability, which we demonstrate on real-world data with up to 300000 variables. Compared to a traditional solver, our approach is over a thousand times faster while maintaining the same solution quality, and in a comparison to the literature, we achieve greater reductions in polarization and disagreement.