Wolt deploys hybrid Transformer-GBDT ranker to boost store trials

Researchers have developed a new hybrid ranking system called the Universal Venue Ranker (UVR) for Wolt, a delivery platform. UVR combines a Transformer encoder for user modeling with a GBDT ranker to integrate various features and enforce local delivery constraints. This unified system replaces four previous ranking models and aims to balance surfacing new stores for trial with maintaining quality for reorder sessions. While UVR increased trial rates by up to 30% in offline tests, it caused regressions in reorder metrics, resulting in statistically unchanged overall conversion rates. However, subsequent A/B tests showed improvements in merchant trial rates and global conversion rates, leading to significant incremental order value and a simplified serving infrastructure. AI

IMPACT This hybrid ranking approach could inform other platforms seeking to balance discovery and retention in recommendation systems.

RANK_REASON Academic paper detailing a new hybrid ranking model for a specific industry application. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• GBDT
• Hugging Face
• Transformer
• Transformer-GBDT Ranker
• Universal Venue Ranker
• Wolt

• product
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →