Calibrated Recommendations with Contextual Bandits

2025-09-05

AI Summary

Spotify’s homepage exhibits severe content-type imbalance—overwhelmingly favoring music due to historical data skew—while user preferences dynamically shift across temporal, weekly, and device-based contexts, resulting in insufficient exposure for underrepresented modalities (e.g., podcasts). To address this, we propose a context-aware online calibration framework grounded in contextual bandits. Unlike conventional approaches relying on static historical averages, our method models real-time user interest distributions across multidimensional contexts to dynamically adjust recommendation weights per content type (music, podcasts, audiobooks). The solution requires no model retraining and integrates seamlessly into existing production systems. Offline evaluations and large-scale A/B tests demonstrate significant improvements: overall user engagement increases substantially, podcast click-through rate rises by 23.6%, and cross-context recommendation diversity and personalization both improve. Critically, the approach effectively mitigates the long-tail content-type bias induced by dataset skew.

Technology Category

• Machine Learning: Online Learning & Bandits
• Data Mining & Knowledge Management: Recommender Systems
• Search and Optimization: Algorithm Configuration

Application Category

• User Modeling, Personalization and Recommendation: Personalized content ranking and presentation
• Web Mining and Content Analysis: Content-based information diffusion
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Spotify's Home page features a variety of content types, including music, podcasts, and audiobooks. However, historical data is heavily skewed toward music, making it challenging to deliver a balanced and personalized content mix. Moreover, users' preference towards different content types may vary depending on the time of day, the day of week, or even the device they use. We propose a calibration method that leverages contextual bandits to dynamically learn each user's optimal content type distribution based on their context and preferences. Unlike traditional calibration methods that rely on historical averages, our approach boosts engagement by adapting to how users interests in different content types varies across contexts. Both offline and online results demonstrate improved precision and user engagement with the Spotify Home page, in particular with under-represented content types such as podcasts.

Problem

• Balancing personalized content mix skewed by historical data
• Adapting to varying user preferences across different contexts
• Boosting engagement with underrepresented content types like podcasts

Innovation

• Contextual bandits for dynamic content distribution
• Personalized calibration based on user context
• Adaptive learning of optimal content type mix