Addressing cold start with dataset transfer in e-commerce learning to rank

Abstraction: Inverse propensity weighting for LTR cold start without historical interactions

Key points:

• Learning-to-rank (LTR) models are key for relevance ranking in e-commerce; training typically requires historical customer interaction logs
• Cold start arises when launching a new platform or extending to new product categories where no interaction history exists
• Proposes using inverse propensity weighting (IPW) as a strategy to create synthetic training datasets when no historical data is available
• Validated via online A/B test demonstrating efficacy of the IPW-based cold-start approach
• Simple and practical enough to be adopted by LTR practitioners without significant infrastructure changes

Connections: Amazon · Learning To Rank · Cold Start Problem · Inverse Propensity Weighting

Source: https://www.amazon.science/publications/addressing-cold-start-with-dataset-transfer-in-e-commerce-learning-to-rank