calibrated recommendations

**Calibrated recommendations** match **user's actual preference distribution** — if a user likes 70% action movies and 30% comedies, recommendations should reflect that ratio, ensuring recommendations align with user's true taste profile rather than over-optimizing for single preferences. **What Is Calibration?** - **Definition**: Recommendations match user's preference distribution. - **Example**: User likes 60% rock, 30% jazz, 10% classical → recommendations should reflect this ratio. - **Goal**: Balanced recommendations reflecting full taste profile. **Why Calibration Matters?** - **User Satisfaction**: Users want variety matching their tastes. - **Avoid Over-Specialization**: Don't only recommend user's #1 preference. - **Fairness**: Give all user interests appropriate attention. - **Discovery**: Maintain exposure to all user interests. - **Long-Term**: Prevent narrowing of user interests over time. **Calibration vs. Accuracy** **Accuracy**: Predict what user will like (may focus on dominant preference). **Calibration**: Match distribution of user's preferences (balanced across interests). **Trade-off**: Most accurate items may not be calibrated. **Measuring Calibration** **KL Divergence**: Distance between user preference distribution and recommendation distribution. **Distribution Matching**: Compare histograms of user preferences vs. recommendations. **Category Coverage**: Ensure all user interest categories represented. **Calibration Techniques** **Re-Ranking**: Adjust recommendation order to match preference distribution. **Sampling**: Sample recommendations from user's preference distribution. **Constraint Optimization**: Optimize accuracy subject to calibration constraints. **Multi-Objective**: Balance accuracy and calibration objectives. **Applications**: Music recommendations (genre diversity), news (topic diversity), e-commerce (product category diversity), video streaming. **Challenges**: Estimating user preference distribution, balancing calibration with accuracy, handling evolving preferences. **Tools**: Calibrated recommendation algorithms, distribution matching methods. Calibrated recommendations provide **balanced, satisfying experiences** — by matching user's full taste profile rather than over-optimizing for dominant preferences, calibration ensures recommendations feel right and maintain user interest diversity.