Active Ranking without Strong Stochastic Transitivity

Authors: Hao Lou, Tao Jin, Yue Wu, Pan Xu, Quanquan Gu, Farzad Farnoud

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Thorough numerical experiments in various settings are conducted, demonstrating that Probe-Rank is significantly more sample-efficient than the state-of-the-art active ranking method. In this section, we present numerical experiments demonstrating the practical performance of Probe Rank.
Researcher Affiliation Academia Hao Lou Dept. of Electrical & Computer Engineering University of Virginia Charlottesville, VA 22903 haolou@virginia.edu Tao Jin Department of Computer Science University of Virginia Charlottesville, VA 22903 taoj@virginia.edu Yue Wu Department of Computer Science University of California, Los Angeles Los Angeles, CA 90095 ywu@cs.ucla.edu Pan Xu Dept. of Biostatistics & Bioinformatics Duke University Durham, NC 27705 pan.xu@duke.edu Quanquan Gu* Department of Computer Science University of California, Los Angeles Los Angeles, CA 90095 qgu@cs.ucla.edu Farzad Farnoud Dept. of Electrical & Computer Engineering University of Virginia Charlottesville, VA 22903 farzad@virginia.edu
Pseudocode Yes Subroutine 1 Successive-Comparison(i, j, δ, τ) (SC) / Algorithm 2 Probe-Rank / Subroutine 3 Probe-Max(S, δ)
Open Source Code Yes Our implementation can be found on Github 5. (Footnote 5: https://github.com/tao-j/aht/releases/tag/v0.1)
Open Datasets No The paper generates synthetic data based on various probabilistic comparison models and random ground truth rankings for its experiments. It does not use or provide concrete access to a pre-existing publicly available dataset. "The probabilistic comparison model pij is generated in different ways to satisfy different assumptions. Note that and d are tuning parameters in all the following settings."
Dataset Splits No The paper describes generating synthetic data for each trial and averaging results over 100 independent trials. It does not refer to traditional train/validation/test splits of a fixed dataset, as its focus is on active learning sample complexity rather than model training on a static dataset.
Hardware Specification Yes We use internal clusters of intel Skylake generation CPUs.
Software Dependencies No The paper mentions that the implementation is on GitHub but does not specify any particular software libraries, frameworks, or their version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes The confidence level δ is fixed to be 0.1. / Specifically, we want to rank n items with the true ranking σ1 σ2 σn, where n varies over [10, 100]. / The probabilistic comparison model pij is generated in different ways to satisfy different assumptions. / The parameter d is set to be 0.3. / We choose α = 1 (see Figure 3(a)) and α = 1/2 (see Figure 3(b)). / In Figure 4, we fix n = 80 and let d vary.