Adaptive Neural Ranking Framework: Toward Maximized Business Goal for Cascade Ranking Systems

Learning to Rank Rank (graph theory) Ranking SVM
DOI: 10.48550/arxiv.2310.10462 Publication Date: 2023-01-01
ABSTRACT
Cascade ranking is widely used for large-scale top-k selection problems in online advertising and recommendation systems, learning-to-rank an important way to optimize the models cascade ranking. Previous works on usually focus letting model learn complete order or order, adopt corresponding rank metrics (e.g. OPA NDCG@k) as optimization targets. However, these targets can not adapt various scenarios with varying data complexities capabilities; existing metric-driven methods such Lambda framework only a rough upper bound of limited metrics, potentially resulting sub-optimal performance misalignment. To address issues, we propose novel perspective optimizing systems by highlighting adaptability capabilities. Concretely, employ multi-task learning adaptively combine relaxed full targets, which refers Recall@m@k respectively. We also introduce permutation matrix represent differentiable sorting techniques relax hard controllable approximate error bound. This enables us both directly more appropriately. named this method Adaptive Neural Ranking Framework (abbreviated ARF). Furthermore, give specific practice under ARF. use NeuralSort obtain draw variant uncertainty weight proposed losses jointly. Experiments total 4 public industrial benchmarks show effectiveness generalization our method, experiment shows that has significant application value.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES ()
CITATIONS ()
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....