PermR algorithm boosts e-commerce revenue by 2% while preserving relevance

A new paper by Svetlana Shirokovskikh and colleagues presents PermR, a permutation-based constrained reranking algorithm designed for e-commerce search and recommendation systems. The core challenge: maximizing revenue from paid promotion products without sacrificing user experience (relevance) or increasing fraud risk. The researchers frame this as an integer linear program (ILP), but solving ILPs per query is too slow for online deployment. PermR sidesteps this by iteratively swapping neighboring items in the ranking, either to improve revenue or to repair a violated constraint. This lightweight approach achieves approximately 63% of the ILP's revenue improvement while keeping all constraints satisfied and staying within production latency requirements.

The algorithm was tested extensively on a large classified platform. In offline experiments across multiple product categories, PermR consistently preserved relevance metrics while boosting revenue. The real-world impact came from a 14-day online A/B test covering over 56 million search queries: PermR drove a 2% revenue increase. This demonstrates that even a modest approximation of the optimal ILP solution can yield significant business value. The work is relevant for any e-commerce platform looking to monetize search results without degrading user trust, offering a practical, deployable alternative to more computationally intensive optimization methods.