Approximate Heavily-Constrained Learning with Lagrange Multiplier Models

Authors: Harikrishna Narasimhan, Andrew Cotter, Yichen Zhou, Serena Wang, Wenshuo Guo

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present experiments on (i) fairness task with intersectional protected groups, (ii) a fairness task with noisy protected groups, and (iii) a ranking fairness task with per-query constraints.
Researcher Affiliation Collaboration Harikrishna Narasimhan, Andrew Cotter, Yichen Zhou Google Research, USA {hnarasimhan,acotter,yichenzhou}@google.com Serena Wang, Wenshuo Guo University of California, Berkeley {serenalwang,wsguo}@berkeley.edu
Pseudocode No The paper describes algorithmic approaches in text (e.g., Section 5.1) but does not provide structured pseudocode or algorithm blocks.
Open Source Code Yes Code available at: https://github.com/google-research/google-research/tree/master/many_constraints.
Open Datasets Yes We use the Communities and Crime dataset [26], which contains 1,994 communities in the US described by 140 features, and seek to predict the per capita crime rate for each community. We use the UCI Adult dataset [26]. We use the Microsoft Learning to Rank Dataset (MSLR-WEB10K) [27].
Dataset Splits No The paper explicitly mentions 'training set' and 'test set' but does not specify a 'validation set' or its split.
Hardware Specification No The paper mentions running times in Appendices D-F but does not explicitly describe the specific hardware (e.g., GPU/CPU models) used for experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We experiment with five multiplier architectures, ranging from under-parameterized to over-parameterized models. This includes a common multiplier for all constraints, a linear model, and neural networks with one, two and three hidden layers (with 50 nodes each). We use a one-hidden layer multiplier neural network model containing 128 nodes to assign a Lagrange multiplier for each query q Q, with the average of the feature vectors within a query used as input. The ranking model is also a one-hidden layer neural network.