🤖 AI Summary
In label-scarce Learning-to-Rank (LTR) tasks on tabular data, gradient-boosted decision trees (GBDTs) dominate deep learning models due to insufficient supervision. Method: This work introduces the first systematic investigation into unsupervised pretraining tailored for tabular ranking—incorporating both autoencoding and contrastive learning—to effectively leverage large-scale unlabeled tabular data, followed by end-to-end fine-tuning. Contribution/Results: Our proposed pretraining paradigm significantly boosts deep ranking model performance, consistently surpassing state-of-the-art GBDT baselines across multiple public and private benchmarks. On cross-domain evaluations, it achieves up to a 38% relative gain in NDCG and demonstrates markedly improved robustness to anomalous samples. This challenges the conventional wisdom that GBDTs are inherently superior under low-label regimes and establishes a new paradigm for deep LTR in data-scarce scenarios.
📝 Abstract
On tabular data, a significant body of literature has shown that current deep learning (DL) models perform at best similarly to Gradient Boosted Decision Trees (GBDTs), while significantly underperforming them on outlier data. However, these works often study idealized problem settings which may fail to capture complexities of real-world scenarios. We identify a natural tabular data setting where DL models can outperform GBDTs: tabular Learning-to-Rank (LTR) under label scarcity. Tabular LTR applications, including search and recommendation, often have an abundance of unlabeled data, and scarce labeled data. We show that DL rankers can utilize unsupervised pretraining to exploit this unlabeled data. In extensive experiments over both public and proprietary datasets, we show that pretrained DL rankers consistently outperform GBDT rankers on ranking metrics -- sometimes by as much as 38% -- both overall and on outliers.
