A Unified Approach to Fair Online Learning via Blackwell Approachability

Authors: Evgenii Chzhen, Christophe Giraud, Gilles Stoltz

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical The goal of this work is to bring to light Blackwell s approachability theory as a suitable theoretical formalism for fair online learning under group fairness constraints. The appealing feature of this theory is two-fold: first, it gives explicit criteria when learning is possible; second, if this criteria is met, it comes with an explicit strategy. ... We apply the general formalism of approachability theory to give new insights into online learning under fairness constraints, and approach this goal in a unified (and geometric) way. In particular, the generality of this formalism allows to derive (im)possibility results nearly effortlessly.
Researcher Affiliation Academia Université Paris-Saclay, CNRS, Laboratoire de mathématiques d Orsay, 91405, Orsay, France {evgenii.chzhen, christophe.giraud, gilles.stoltz} @universite-paris-saclay.fr
Pseudocode Yes PROTOCOL 2.1 Parameters: Observation operator G for Nature; distribution Q on X S For t = 1, 2, . . . 1. Contexts (xt, st) are sampled according to Q, independently from the past; 2. Simultaneously, Nature observes G(xt, st) and picks bt B; the Player observes xt and picks an action at A; 3. The Player observes the reward m(at, bt, xt, st) and the sensitive context st, while Nature observes (at, bt, xt, st).
Open Source Code No The paper is theoretical and does not involve experimental code release. The authors indicate 'N/A' for questions about code in the 'Ethics Review Checklist': '3. If you ran experiments... (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [N/A]'
Open Datasets No The paper is theoretical and does not conduct experiments involving actual datasets. It discusses theoretical constructs like 'stochastic sensitive and non-sensitive contexts' and 'context distribution Q', but these are not actual datasets requiring public access information.
Dataset Splits No The paper is theoretical and does not conduct experiments involving data splits (training, validation, test sets). Therefore, no information on dataset splits is provided.
Hardware Specification No The paper is theoretical and does not describe any computational experiments or hardware used. The authors indicate 'N/A' for questions about hardware in the 'Ethics Review Checklist': '3. If you ran experiments... (d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [N/A]'
Software Dependencies No The paper is theoretical and does not describe any computational experiments or software dependencies with specific version numbers.
Experiment Setup No The paper is theoretical and does not describe an experimental setup, hyperparameters, or training details for empirical evaluation. The authors indicate 'N/A' for questions about training details in the 'Ethics Review Checklist': '3. If you ran experiments... (b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [N/A]'