How does overparametrization affect performance on minority groups?

Authors: Saptarshi Roy, Subha Maity, Songkai Xue, Mikhail Yurochkin, Yuekai Sun

TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this paper, we complement these empirical studies with a theoretical investigation of the risk of overparameterized random feature regression models on minority groups... In an experiment with California housing prices dataset 2 (see Appendix D)... we also provide a simulation study for the effect of overparameterization on for classifications with random features.
Researcher Affiliation Collaboration Saptarshi Roy EMAIL Department of Computer Science University of Texas at Austin; Subha Maity EMAIL Department of Statistics & Actuarial Science University of Waterloo; Songkai Xue EMAIL Department of Statistics University of Michigan, Ann Arbor; Mikhail Yurochkin EMAIL MIT-IBM Watson AI Lab; Yuekai Sun EMAIL Department of Statistics University of Michigan, Ann Arbor
Pseudocode No The paper does not contain any explicitly labeled pseudocode or algorithm blocks. It primarily presents mathematical derivations and theoretical analyses.
Open Source Code Yes 1Codes are available at https://github.com/smaityumich/overparameterized-group-fairness.
Open Datasets Yes In an experiment with California housing prices dataset 2 (see Appendix D)... 2https://www.kaggle.com/datasets/camnugent/california-housing-prices
Dataset Splits Yes Furthermore, we split the data into training (80%) and test (20%) datasets.
Hardware Specification No The paper does not provide specific details about the hardware used for running its experiments or simulations.
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes In the simulation for Figure 8 we let σ( ) be the Re LU activation function and θi,j be IID standard normal distributed. Moreover, we let π = 0.95, β0 = 10e1, β1 = 10 cos(θ)e1 + 10 sin(θ)e2, n = 400, d = 200, N = γn where e1 and e2 are the first two standard basis of Rd. We tune hyperparameters θ {0 , 45 , 90 , 135 , 180 } and γ {0.5, 1, . . . , 3}, then report test errors averaged over 20 replicates.