CSTS algorithm balances multiple goals for public media recommendations

Public service media recommender systems face the challenge of balancing competing objectives such as audience engagement, cultural diversity, and operational constraints. Existing methods often use fixed weight combinations or Pareto optimization, which fail to adapt when priorities shift across different situations. To address this, Théo Maëtz, Luc Guillet, and Andrea Cavallaro introduce the Contextual Scalarisation Thompson Sampler (CSTS), a multi-objective contextual bandit algorithm that learns to adjust the importance of each objective based on the observed context.

CSTS was evaluated using real programming data from Radio Télévision Suisse (RTS), the Swiss national broadcaster. The results showed that CSTS significantly improved contextual relevance and better mimicked expert curation decisions compared to both fixed-weight scalarization and standard contextual bandit baselines. The paper, accepted at ICPR 2026, includes 15 pages, 3 figures, and 3 tables. This approach offers a practical solution for any organization needing to balance multiple, sometimes conflicting, goals in dynamic recommendation scenarios.