Multileave Gradient Descent for Fast Online Learning to Rank

Summary

This research presents Multileave Gradient Descent (MGD), an extension of Dueling Bandit Gradient Descent for online learning to rank in search engines. The approach enables exploration of multiple ranking directions simultaneously before updating, rather than exploring a single direction at a time, using multileaved comparisons instead of expensive pairwise interleaving comparisons. Experimental validation demonstrates that MGD achieves large improvements over existing methods by reducing the number of updates needed while maintaining effectiveness in learning from user interactions.