Slate Bandit

The foundational framework from which slate bandits extend. A contextual bandit observes a feature vector (user profile, session state) before selecting a single action. The slate variant generalizes this to select an ordered set of K items simultaneously. Key distinction: standard bandits receive feedback on one arm, while slate bandits receive feedback on the entire composed list, requiring specialized reward decomposition techniques.

A specific slate bandit model that assumes users scan items top-to-bottom and click on the first attractive item, ignoring the rest. The cascade model simplifies reward attribution: each item's reward is conditional on all higher-ranked items being skipped. This assumption reduces computational complexity and is widely used in search result ranking where position bias is dominant.

A critical evaluation technique for slate recommenders. When training on logged data from a production system, the probability of observing a specific ordered slate under the logging policy must be computed. Standard Inverse Propensity Scoring (IPS) becomes intractable for large slates due to combinatorial explosion. Specialized estimators use importance sampling with pseudo-likelihood approximations to debias feedback.

A probabilistic model for ensuring diversity within a slate. A DPP defines a distribution over subsets where the probability of selecting a set is proportional to the determinant of a kernel matrix. Items with high similarity have low co-selection probability. In slate bandits, DPPs balance relevance with coverage, preventing the algorithm from recommending near-duplicate items in the same list.

The central technical challenge of slate bandits. The environment returns a single slate-level reward (e.g., session engagement), but the algorithm must assign credit to individual items for learning. Approaches include:

• Additive decomposition: Assume total reward is sum of individual item rewards
• Cascade model: Attribute reward to the first engaged item
• Shapley values: Game-theoretic marginal contribution of each item Incorrect decomposition leads to biased learning.

A statistical framework for modeling how an item's rank position within a slate affects its probability of engagement, independent of relevance. The examination hypothesis states: P(click) = P(examination | position) × P(relevance | item, context). Slate bandits must learn position bias parameters jointly with item relevance to avoid conflating high-ranked items with truly preferred items.