Feature Noise Induces Loss Discrepancy Across Groups

Authors: Fereshte Khani, Percy Liang

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our results on three real-world datasets: for predicting the final grade of secondary school students , final GPA of law students , and crime rates in the US communities , where the group g is either race or gender.
Researcher Affiliation Academia Fereshte Khani 1 Percy Liang 1 1Department of Computer Scinece, Stanford University. Correspondence to: Fereshte Khani <fereshte@stanford.edu>.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Reproducibility. All code, data and experiments for this paper are available on the Coda Lab platform at https: //worksheets.codalab.org/worksheets/ 0x7c3fb3bf981646c9bc11c538e881f37e.
Open Datasets Yes We consider three real-world datasets from the fairness literature. See Table 2 for a summary and Appendix G for more details. ... for predicting the final grade of secondary school students (Cortez and Silva, 2008), final GPA of law students (Wightman and Ramsey, 1998), and crime rates in the US communities (Redmond and Baveja, 2002).
Dataset Splits No We run each experiment 100 times, each time randomly performing a 80 20 train-test split of the data, and reporting the average on the test set. This specifies only a train-test split, without explicit mention of a separate validation split.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU, CPU models, memory) used to run the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., library names with specific versions) required to replicate the experiments.
Experiment Setup Yes We standardize all features and the target in all datasets (except the group membership feature) to have mean 0 and variance 1. We run each experiment 100 times, each time randomly performing a 80 20 train-test split of the data, and reporting the average on the test set. We compute the least squares estimator for each of the two observation functions: o g which only have access to non-group features, and o+g which have access to all features. We consider two types of noise: 1. Equal noise: for different values of σ2 u we add independent normal noise (u N(0, σ2 u)) to each feature except the group membership. 2. Omitting features: We start with a random order of the non-group features and omit features, which is nearly equivalent to adding normal noise with a very high variance (u N(0, 10000)) to them sequentially.