Learning Cascade Ranking as One Network

Authors: Yunli Wang, Zhen Zhang, Zhiqiang Wang, Zixuan Yang, Yu Li, Jian Yang, Shiyang Wen, Peng Jiang, Kun Gai

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that LCRON achieves significant improvement over existing methods on public benchmarks and industrial applications, addressing key limitations in cascade ranking training and significantly enhancing system performance.
Researcher Affiliation Collaboration 1Kuaishou Technology, Beijing, China 2Beihang University, Beijing, China 3Independent, Beijing, China. Correspondence to: Yunli Wang, Jian Yang <EMAIL, EMAIL>.
Pseudocode No The paper describes its methodology using mathematical formulations and descriptive text in Section 4. 'Methodology' but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The source code of our public experiments is publicly available2. 2https://github.com/Kwai/LCRON
Open Datasets Yes We conduct public experiments based on Rec Flow (Liu et al., 2025), which, to the best of our knowledge, is the only public benchmark that collects data from all stages of real-world cascade ranking systems.
Dataset Splits Yes Following the mainstream setup for evaluating recommendation datasets, we use the last day of Period 1 as the test set to report the main results of our experiments (Section 5.3), while the second-to-last day serves as the validation set for tuning the hyperparameters... In this setting, when day t is designated as the test set, the corresponding training data encompass all days from the beginning of Period 1 up to day t 1.
Hardware Specification Yes To validate this, we conducted experiments on Rec Flow using A800 GPUs and recorded the GPU memory usage and runtime.
Software Dependencies Yes All offline experiments are implemented using Py Torch 1.13 in Python 3.7.
Experiment Setup Yes We employ the Adam optimizer with a learning rate of 0.01 for training all methods. Following the common practice in online recommendation systems (Liu et al., 2025; Zhang et al., 2022), each method is trained for only one epoch. The batch size is set to 1024. The source code of our public experiments is publicly available2. We tune the hyper-parameter τ on the validation set, which controls the temperature of Neural Sort. We set q1 and q2 in Le2e to 10 during training.