Robust Consensus in Ranking Data Analysis: Definitions, Properties and Computational Issues

Authors: Morgane Goibert, Clément Calauzènes, Ekhine Irurozki, Stephan Clémençon

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Beyond the theoretical contributions, the relevance of the approach proposed is supported by an experimental study. [...] In this section, we illustrate the relevance of the statistic outputted by our Downward Merge plug-in on Kemeny s median (called our Downward Merge statistic for short) by running several illustrative experiments for various settings and comparing with the baseline provided by the usual Kemeny s median.
Researcher Affiliation Collaboration 1Criteo AI Lab, Paris, France 2T el ecom Paris, Paris, France. Correspondence to: Morgane Goibert <morgane.goibert@gmail.com>.
Pseudocode Yes Algorithm 1 Na ıve Merge [...] Algorithm 2 Downward Merge
Open Source Code Yes The code is available here.
Open Datasets Yes To corroborate these findings, we also ran experiments using real-world datasets from the preflib library: two Netflix Prize datasets (resp. with n = 3 and n = 4 items), a Debian dataset (with n = 5 items) and an Apa dataset (with n = 5 items).
Dataset Splits No The paper mentions using "various distributions p" and "real-world datasets" but does not specify any training, validation, or test dataset splits, percentages, or procedures for partitioning the data.
Hardware Specification No The paper does not provide any details about the hardware (e.g., GPU/CPU models, memory, or specific computing environments) used to run the experiments.
Software Dependencies No The paper mentions using the "preflib library" but does not specify any software names with version numbers (e.g., Python version, specific library versions, or frameworks like PyTorch or TensorFlow).
Experiment Setup Yes When the threshold is set to a sensible value (here θ = 0.05), the Downward Merge algorithm outputs a bucket order as a statistic: thus, the robustness increases very strongly to reach nearly optimal values even for very small values of δ, which illustrates its efficiency. [...] Figure 5 shows the results for different choices of distribution p when the number of items n = 4, and for δ = 1/6 (normalized value of δ that requires at least a switch between two items to break the statistic).