Certifying Robustness to Programmable Data Bias in Decision Trees
Authors: Anna Meyer, Aws Albarghouthi, Loris D'Antoni
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach on datasets that are commonly used in the fairness literature, and demonstrate our approach s viability on a range of bias models. We evaluate our approach on a number of bias models and datasets from the fairness literature. Our tool can certify pointwise robustness for a variety of bias models; we also show that some datasets have unequal robustness-certiļ¬cation rates across demographics groups. |
| Researcher Affiliation | Academia | Anna P. Meyer, Aws Albarghouthi , and Loris D Antoni Department of Computer Sciences University of Wisconsin Madison Madison, WI 53706 {annameyer, aws, loris}@cs.wisc.edu |
| Pseudocode | No | The paper describes the learning algorithm and abstract transformers mathematically and textually but does not provide a formal pseudocode block or algorithm listing. |
| Open Source Code | Yes | For all datasets, we use the standard train/test split if one is provided; otherwise, we create our own train/test splits, which are available in our code repository at https://github.com/annapmeyer/antidote-P. |
| Open Datasets | Yes | We evaluate on Adult Income [17] (training n=32,561), COMPAS [29] (n=4629), and Drug Consumption [20] (n=1262). A fourth dataset, MNIST 1/7 (n=13,007), is in the Appendix. |
| Dataset Splits | No | The paper mentions using "standard train/test split" or creating "our own train/test splits" and points to the code repository for details. It does not explicitly state validation split percentages or methodology within the main text. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper states that the technique is implemented in C++ but does not provide specific version numbers for any software dependencies, libraries, or compilers used. |
| Experiment Setup | Yes | For each dataset, we choose the smallest tree depth where accuracy improves no more than 1% at the next-highest depth. For Adult Income and MNIST 1/7, this threshold is depth 2 (accuracy 83% and 97%, respectively); for COMPAS and Drug Consumption it is depth 1 (accuracy 64% and 76%, respectively). |