PairRank: Online Pairwise Learning to Rank by Divide-and-Conquer

TL;DR

PairRank introduces a divide-and-conquer online learning to rank method that focuses exploration on uncertain pairs, improving empirical performance and theoretical convergence in online ranking tasks.

Contribution

It proposes a novel online pairwise learning to rank approach that partitions candidate documents and targets uncertain pairs, bridging theoretical guarantees with practical effectiveness.

Findings

Outperforms existing OL2R baselines on benchmark datasets

Proven regret bounds linked to ranking accuracy

Effective focus on uncertain pairs enhances learning efficiency

Abstract

Online Learning to Rank (OL2R) eliminates the need of explicit relevance annotation by directly optimizing the rankers from their interactions with users. However, the required exploration drives it away from successful practices in offline learning to rank, which limits OL2R's empirical performance and practical applicability. In this work, we propose to estimate a pairwise learning to rank model online. In each round, candidate documents are partitioned and ranked according to the model's confidence on the estimated pairwise rank order, and exploration is only performed on the uncertain pairs of documents, i.e., \emph{divide-and-conquer}. Regret directly defined on the number of mis-ordered pairs is proven, which connects the online solution's theoretical convergence with its expected ranking performance. Comparisons against an extensive list of OL2R baselines on two public learning to rank benchmark datasets demonstrate the effectiveness of the proposed solution.