Fairness constraints can help exact inference in structured prediction
Authors: Kevin Bello, Jean Honorio
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we corroborate our theoretical results through synthetic experiments. Graphs with high expansion properties such as complete graphs and d-regular expanders are known to manifest high probability of exact recovery as their Cheeger constant increases with respect to n or d (Bello & Honorio 2019). That is, in these graphs, the effect of the fairness constraint will not be noticeable. In contrast, graphs with poor expansion properties such as grids, which have a Cheeger constant in the order of O(1/n) for a Grid(n, n), can only be recovered approximately (Globerson et al. 2015), or exactly if the graph can be perturbed with additional edges (Bello & Honorio 2019). Thus, we focus our experiments on grids and empirically show how the inclusion of the fairness constraint boosts the probability of exact recovery. In Figure 2, we first randomly set y by independently sampling each yi from a Rademacher distribution. We consider a graph of 64 nodes, specifically, Grid(4, 16), i.e., is guaranteed to be greater than 0. Finally, we compute 30 observations for p [0, 0.1]. When there is no fairness constraint, we observe that the probability of exact recovery decreases at a very high rate, while the addition of fairness constraints improves the exact recovery probability. |
| Researcher Affiliation | Academia | Kevin Bello Department of Computer Science Purdue Univeristy West Lafayette, IN 47906, USA kbellome@purdue.edu Jean Honorio Department of Computer Science Purdue Univeristy West Lafayette, IN 47906, USA jhonorio@purdue.edu |
| Pseudocode | No | The paper includes mathematical derivations and proofs but does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not contain any statement about making its source code publicly available, nor does it provide a link to a code repository. |
| Open Datasets | No | The paper describes generating synthetic data for its experiments: "In Figure 2, we first randomly set y by independently sampling each yi from a Rademacher distribution. We consider a graph of 64 nodes, specifically, Grid(4, 16)". It does not provide access information, a citation, or a repository for a pre-existing or publicly released dataset. |
| Dataset Splits | No | The paper describes generating synthetic data and running experiments with "30 observations for p [0, 0.1]" on a "Grid(4, 16)" graph, but it does not specify any train, validation, or test dataset splits, nor does it mention cross-validation or other data partitioning methods. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments, such as GPU/CPU models or other computing resources. |
| Software Dependencies | No | The paper does not specify any software dependencies with version numbers, which are necessary for reproducibility. |
| Experiment Setup | No | The paper states details for its synthetic experiments such as randomly setting 'y' from a Rademacher distribution and using a 'Grid(4, 16)' graph with '30 observations for p [0, 0.1]'. It also mentions how the attribute 'a1' was sampled. However, it does not provide explicit hyperparameters or system-level training settings commonly associated with experimental setups in machine learning, such as learning rates, batch sizes, or optimizer configurations. |