Group Meritocratic Fairness in Linear Contextual Bandits
Authors: Riccardo Grazzi, Arya Akhavan, John IF Falk, Leonardo Cella, Massimiliano Pontil
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We use simulated settings and experiments on the US census data to show that our policy achieves sub-linear fair pseudo-regret also in practice. [...] In Sec. 6 we present an illustrative simulation experiment with diverse reward distributions. In Sec. 7, we extend our policy and results to the case where candidates from the same arm can belong to different groups and show the efficacy of our approach with an experiment on the US census data where the sensitive group (ethnicity) is drawn at random together with the context. |
| Researcher Affiliation | Academia | Riccardo Grazzi1,2 , Arya Akhavan1,3, John Isak Texas Falk1,2, Leonardo Cella1, Massimiliano Pontil1,2 1CSML, Istituto Italiano di Tecnologia, Genoa, Italy 2Dept. of Computer Science, University College London, UK 3CREST, ENSAE, Institut Polytechnique de Paris, France |
| Pseudocode | Yes | Algorithm 1 Fair-Greedy |
| Open Source Code | Yes | Code at https://github.com/CSML-IIT-UCL/GMFbandits |
| Open Datasets | Yes | Finally, we use simulated settings and experiments on the US census data to show that our policy achieves sub-linear fair pseudo-regret also in practice. [...] US Census experiments. Group = Ethnicity. We test this setting in practice by simulating the hiring scenario discussed above with data from the US Census containing the income and other useful indicators of several individuals in the United States. This data is accessed via the Folk Tables library [13]. |
| Dataset Splits | No | The paper does not explicitly mention training/validation/test splits, only refers to "simulated settings" and "US census data". It uses a number of rounds for evaluation (e.g. T=2500 in Figure 2). |
| Hardware Specification | No | The paper does not specify any hardware details like GPU/CPU models, memory, or cloud resources used for experiments. |
| Software Dependencies | No | The paper does not explicitly list software dependencies with version numbers. |
| Experiment Setup | Yes | We set K = 4, ηt = 2ξt, where ξt has standard normal distribution, Xa = Ba Ya + ca where each coordinate of Ya R4 is an independent sample from the uniform distribution on [0, 1], Ba R(4K+1) 4 is such that Xa contains Ya starting from the 4a-th coordinate and ca has all the coordinates set to zero except for the last which is set to 3a to simulate a group bias. In this setup µ acts differently on each group, in particular, we note that µ R4K+1 has its last coordinate multiplying the group bias in ca, which we set to 1, and 4 group-specific coordinates, which we set to manually picked values between 0 and 9. Results are shown in Fig. 1, where we compare our greedy policy in Alg. 1 with OFUL [1], both with regularization parameter set to 0.1, and with the Uniform Random policy. |