Feature Importance Disparities for Data Bias Investigations

Authors: Peter W Chang, Leor Fishman, Seth Neel

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

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we conduct a thorough empirical evaluation, auditing for large FID subgroups on the Student (Cortez & Silva, 2008), COMPAS (Angwin et al., 2016), Bank (Moro et al., 2014), and Folktables (Ding et al., 2021) datasets, using LIME, SHAP, saliency maps, and linear regression coefficient as feature importance notions.
Researcher Affiliation Academia 1SAFR AI Lab, Harvard Business School & Kempner Institute, Boston, MA 2Harvard College, Cambridge, MA. Correspondence to: Seth Neel <sethneel.ai@gmail.com>.
Pseudocode Yes Algorithm 1 Iterative Constrained Optimization
Open Source Code Yes The code used for our experiments is available at github.com/safr-ai-lab/xai-disparity.
Open Datasets Yes We used four popular datasets for the experiments: Student (Cortez & Silva, 2008), COMPAS (Angwin et al., 2016), Bank (Moro et al., 2014), and Folktables (Ding et al., 2021).
Dataset Splits No Datasets were split into 80 20 train-test split except for Student which was split 50 50 due to its small size.
Hardware Specification Yes The optimization was run using GPU computing on NVIDIA Tesla V100s.
Software Dependencies No The parameter vector θ for a logistic regression classifier was randomly initialized with a Py Torch random seed of 0 for reproducability. We used an ADAM (Kingma & Ba, 2015) optimizer with a learning rate of .05 as our heuristic solver for the loss function.
Experiment Setup Yes We empirically found that setting the hyperparameter B = 10^4 µ(fj) worked well on all datasets... We set the learning rate for exponentiated gradient descent to η = 10^-5... We found that setting the error tolerance hyperparameter ν = .05 µ(fj) n αL worked well... For all datasets and methods we ran for at most T = 5000 iterations... We used an ADAM (Kingma & Ba, 2015) optimizer with a learning rate of .05... Empirical testing found that values of λs = 10^5 and λc = 10^-1 returned appropriate subgroup sizes...