Contextual Bandit
A machine learning framework that makes personalization decisions by balancing exploitation of known preferences with exploration of uncertain options, using contextual features about the user and situation to optimize actions.
Contextual bandits extend the classic multi-armed bandit problem by incorporating contextual information about the user and situation when deciding which action to take. For each decision point, the system observes context features, selects an action from available options, and receives a reward signal. Over time, it learns which actions work best in which contexts while continuing to explore uncertain options.
For growth teams, contextual bandits are ideal for personalization problems where you need to continuously learn and adapt: which homepage layout to show, which email subject line to use, which onboarding flow to present, or which product to feature. Unlike A/B tests that run for a fixed period, bandits continuously shift traffic toward better-performing variants while maintaining exploration. AI-powered bandit systems can handle thousands of contextual features and hundreds of actions simultaneously. Growth engineers should implement contextual bandits for decisions that are made frequently, have clear reward signals, and benefit from personalization. The key advantage over traditional A/B testing is efficiency, as bandits minimize regret by reducing exposure to underperforming variants faster, and they naturally handle the personalization case where the best option varies by user context.
Related Terms
Recommendation Engine
A system that uses algorithms and machine learning to suggest relevant items, content, or actions to users based on their behavior, preferences, and similarities to other users, driving engagement and conversion.
Collaborative Filtering
A recommendation technique that predicts a user's preferences by analyzing behavior patterns across many users, based on the principle that people who agreed in the past tend to agree in the future.
Content-Based Filtering
A recommendation approach that suggests items similar to those a user has previously liked or interacted with, based on item attributes and features rather than the behavior of other users.
Matrix Factorization
A mathematical technique used in recommendation systems that decomposes the large, sparse user-item interaction matrix into lower-dimensional latent factor matrices, revealing hidden patterns that predict user preferences.
Cold-Start Problem
The challenge of providing relevant recommendations or personalized experiences to new users with no interaction history or for new items with no engagement data, a fundamental limitation of data-driven personalization systems.
Popularity Bias
The tendency of recommendation systems to disproportionately suggest already popular items, creating a feedback loop where popular items get more exposure and engagement, further reinforcing their dominance over niche content.