Classification Under Misspecification: Halfspaces, Generalized Linear Models, and Evolvability

Authors: Sitan Chen, Frederic Koehler, Ankur Moitra, Morris Yau

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

Reproducibility Variable Result LLM Response
Research Type Experimental Lastly, we empirically evaluate our algorithm for Massart halfspaces and find it exhibits some intriguing fairness properties.
Researcher Affiliation Academia MIT sitanc@mit.edu Frederic Koehler MIT fkoehler@mit.edu Ankur Moitra MIT moitra@mit.edu Morris Yau UC Berkeley morrisyau@berkeley.edu
Pseudocode Yes Algorithm 1: FINDDESCENTDIRECTION(w, ε, δ, λ) and Algorithm 2: FILTERTRON(ε, η, δ, λ, T)
Open Source Code Yes All code for the experiments can be found at https://github.com/secanth/massart.
Open Datasets Yes We evaluated FILTERTRON, gradient descent on the Leaky Relu loss, logistic regression, and random forest (to compare with a less interpretable, non-halfspace classifier) on the UCI Adult dataset, obtained from the UCI Machine Learning Repository [DG17] and originally curated by [Koh96].
Dataset Splits Yes For every p, we took a five-fold cross validation of the dataset, and for every ε ∈ [0, 0.1, 0.2, 0.3, 0.4] we repeated the following five times and took the mean: (1) randomly flip the labels for the training and test set according to the Massart adversary, (2) train on the noisy training set, and (3) evaluate according to (A) and (B).
Hardware Specification Yes The experiments on the Adult dataset were conducted in a Kaggle kernel with a Tesla P100 GPU
Software Dependencies No The paper does not provide specific software versions for its dependencies.
Experiment Setup Yes For FILTERTRON and gradient descent on the Leaky Relu loss, we ran for 2000 iterations with step size 0.05 and ε chosen by a naive grid search over [0.05, 0.1, 0.15, 0.2].