SureMap: Simultaneous mean estimation for single-task and multi-task disaggregated evaluation

Authors: Misha Khodak, Lester Mackey, Alexandra Chouldechova, Miro Dudik

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Sure Map on disaggregated evaluation tasks in multiple domains, observing significant accuracy improvements over several strong competitors.
Researcher Affiliation Collaboration Mikhail Khodak Princeton University mkhodak@cs.cmu.edu Lester Mackey, Alexandra Chouldechova, Miroslav Dudík Microsoft Research {lmackey,alexandrac,mdudik}@microsoft.com
Pseudocode Yes Algorithm 1: Single-task Sure Map. (For multi-task Sure Map see D.)
Open Source Code Yes Code for both generating the task data and reproducing the method evaluations is available at https://github.com/mkhodak/Sure Map.
Open Datasets Yes Diabetes. This is a tabular dataset of Strack et al. [2014]... Adult. We use the classic Adult census dataset [Kohavi, 1996]... State-Level ACS (SLACS). ... assembled by Ding et al. [2021]... Common Voice (CV) dataset [Ardila et al., 2020]
Dataset Splits Yes Common Voice. This is a single-task dataset obtained by combining the validation and test partitions of the CV dataset.
Hardware Specification Yes By far the most computation was required to generate the Common Voice, CVC, and Adult tasks, which was done on a machine with two RTX-8000 GPUs and took about a week.
Software Dependencies No The paper mentions software like 'Whisper ASR model', 'llama-3-70b', and 'Sci Py implementation' but does not provide specific version numbers for these or other key software dependencies.
Experiment Setup Yes Our main metric is MAE relative to a ground truth vector, which we take to be the mean of all available data for each subpopulation g [d], except those with fewer than 40 samples. In our main results we subsample with replacement from the entire dataset at different rates and track performance as a function of the sizes of the resulting datasets. To obtain 95% confidence intervals we conduct 200 and 40 random trials at each subsampling rate in the single-task and multi-task settings, respectively.