Reproduce “PiRank: Scalable Learning To Rank via Differentiable Sorting” on OpenResearch

About this paper

Robin Swezey, Aditya Grover, Bruno Charron, Stefano Ermon

PiRank introduces a scalable learning-to-rank framework by employing differentiable sorting relaxations coupled with a divide-and-conquer strategy to directly optimize non-differentiable ranking metrics. The method significantly reduces computational complexity for large item lists and demonstrates improved performance on public LTR benchmarks compared to existing baselines.

Key insights

• Differentiable relaxations of sorting allow for principled, direct optimization of non-differentiable ranking metrics, establishing a closer link between training objectives and evaluation metrics.
• A divide-and-conquer approach can effectively overcome the scalability limitations of differentiable sorting, making listwise LTR practical for extensive item lists.
• The temperature parameter in the differentiable relaxation is critical for training stability and convergence, balancing gradient smoothness and bias.
• Training with larger list sizes (L_train) generally improves NDCG@k performance, particularly when the test list size (L_test) is large and the focus is on a small k.

Summary by alphaXiv.