Contextual Scalarisation Thompson Sampling for multi-objective decisions in public media

Technical Summary: Contextual Scalarisation Thompson Sampling for Multi-Objective Decisions in Public Media

Problem Statement
Public service media organizations face a complex multi-objective optimization problem when scheduling programming. Unlike commercial systems that often prioritize short-term engagement, public broadcasters must balance competing criteria such as audience reach, cultural values, diversity, local relevance, minority representation, and operational constraints (e.g., rights management and budget). Traditional approaches rely on fixed combinations of objectives or Pareto-based optimization, which fail to adapt to changing priorities across different broadcasting contexts (e.g., time slots, seasons, or competing events). Furthermore, existing scheduling tools often lack the ability to model editorial tradeoffs dynamically, leaving curators to manually weigh factors for each decision, a process that does not scale well with large catalogues.

Methodology
The authors propose Contextual Scalarisation Thompson Sampling (CSTS), a multi-objective contextual bandit algorithm designed to learn context-dependent weights for multiple value signals.

• Framework: CSTS formulates the scheduling problem as a contextual multi-armed bandit. At each time step $t$, given a context $x_{c,t}$ (describing the slot, channel, season, and competition) and a set of candidate actions $A_t$ (available films), the system computes a vector of $N$ bounded value signals $\phi(a_t, x_{c,t})$. In this study, $N=5$, representing:

  1. Audience potential
  2. Diversity
  3. Novelty
  4. Competition (concurrent airing programs)
  5. Rights/Contract management

• Adaptive Scalarisation: Instead of using fixed global weights, CSTS learns a weighting function $w(x_{c,t})$ that maps the context to a probability vector over the objectives. This is parametrized as a linear gating model passed through a softmax function:
  $$w_\theta(x_{c,t}) = \text{softmax}(U x_{c,t} + b)$$
  The scalar utility for an action is then calculated as the dot product of these context-dependent weights and the value signal vector: $u(a_t, x_{c,t}) = w_\theta(x_{c,t})^\top \phi(a_t, x_{c,t})$.

• Exploration Strategy: To balance exploration and exploitation, the algorithm employs Thompson Sampling on the model parameters $\theta = (U, b)$. At each step, it samples parameters $\tilde{\theta}_t$ from a distribution centered on the current estimate (with uncertainty modeled via an RMS-type accumulator). This allows the system to explore different plausible tradeoffs between objectives rather than injecting arbitrary noise into item scores.

• Learning and Feedback: The system updates parameters based on binary feedback (curator selection) using a logistic loss function. Additionally, it incorporates an auxiliary loss to align the learned weights with explicit curator feedback regarding specific signal priorities, effectively steering the model toward editorial intent.

Key Contributions

1. Context-Dependent Weighting: Unlike prior work that relies on fixed weights or static Pareto frontiers, CSTS dynamically adjusts the relative importance of objectives based on the observed context.
2. Interpretability and Efficiency: By decomposing recommendation quality into bounded value signals and learning a context-dependent weight vector, the method retains the data efficiency and interpretability of scalar reward bandits while handling multiple objectives.
3. Decision Support System: The approach is designed specifically to assist human curators by surfacing a ranked shortlist of plausible candidates that align with expert curation practices, rather than fully automating the schedule.

Results
The authors evaluated CSTS using two years of historical programming data from Radio Télévision Suisse (RTS), comparing it against static global weights, audience maximization (single-objective), vanilla Thompson Sampling, and LinUCB.

• Contextual Relevance: CSTS achieved the highest relaxed contextual relevance (Hit@10 = 98.7%, NDCG@10 = 0.376), significantly outperforming vanilla Thompson Sampling (80.0%) and static weights (92.0%). This demonstrates the benefit of adapting weights to the specific context.
• Strict Matching: Under strict matching (where relevance is defined solely as the historically chosen film), static global weights performed slightly better (Hit@10 = 20.0% vs. 18.7% for CSTS). The authors attribute this to the fact that some historical choices were punctual reactions to last-minute constraints not fully captured in the context features, and that expert-defined global weights serve as a strong baseline.
• Tradeoff Analysis: CSTS showed a distinct preference profile compared to static methods. While static recommendations consistently favored Novelty, CSTS learned to prioritize Diversity and Rights urgency in 3 out of 4 tested contexts, aligning more closely with catalogue efficiency and rights management.
• Ablation Studies: Removing any single value signal degraded performance, confirming the necessity of the multi-objective approach. The removal of "Competition" or "Audience" signals caused the largest drops in relevance. Exploration tuning (via the parameter $\kappa$) showed that moderate exploration yields the best accuracy, preventing overfitting while maintaining coverage.

Significance and Claims
The paper claims that CSTS offers a significant improvement in contextual relevance and alignment with expert curation practices compared to fixed-weight or standard contextual bandit approaches. It successfully demonstrates that a multi-objective bandit can learn to navigate the tradeoffs inherent in public service media scheduling without requiring a static, one-size-fits-all policy.

The authors modestly note that while CSTS improves the retrieval of relevant slates, it does not yet fully replicate the exact historical choices of curators, suggesting that human decision-making involves factors beyond the captured context features. They emphasize that the system is intended as a decision support tool to help curators navigate large catalogues, rather than a replacement for human editorial judgment. Future work is proposed to move from hand-crafted signals to a more systematic design process and to integrate digital audience measurements.