Personalizing Carousel Ranking on Wolt's Discovery Page: A Hierarchical Multi-Armed Bandit Approach

Wolt is personalizing carousel ranking on their discovery page using a hierarchical multi-armed bandit approach to improve conversion rates and user experience

Key architectural requirements included:

• Real-time serving capabilities
• High availability
• Easy maintenance
• Integration with existing ML platform
• Regular data updates without redeployment

Implementation uses a two-level hierarchical approach:

• City level provides prior information
• User level personalizes based on purchase history
• Updates happen daily to incorporate new user data

Main data signals used:

• Purchase history as primary signal
• Time of day for context awareness
• City-wide conversion rates as prior information
• Impression tracking for position bias

Technical implementation details:

• Snowflake for data storage/processing
• Flight for ML pipelines
• Kubernetes for deployment
• Feature store for serving
• A/B testing across three user groups

Key challenges addressed:

• Cold start problems
• Position bias
• Multiple stakeholder requirements
• Real-time latency constraints
• Data quality and validation

Future improvements planned:

• More sophisticated exploration mechanisms
• Multi-objective optimization
• Enhanced context awareness
• Better diversity handling
• Dynamic updates

Results show personalization matters:

• Improved conversion rates
• Better user engagement
• More relevant carousel ordering
• Successful handling of different user intents and contexts