Selective Regression under Fairness Criteria
Authors: Abhin Shah, Yuheng Bu, Joshua K Lee, Subhro Das, Rameswar Panda, Prasanna Sattigeri, Gregory W Wornell
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The effectiveness of these approaches is demonstrated on synthetic and real-world datasets. Finally, we empirically1 demonstrate the effectiveness of these algorithms on real-world datasets (see Section 6). We test our algorithms on Insurance and Crime datasets, and provide an application of our method in Causal Inference via IHDP dataset. |
| Researcher Affiliation | Collaboration | 1Massachusetts Institute of Technology 2This work was done while J. Lee was at Massachusetts Institute of Technology; the author is now with Snap 3MIT-IBM Watson AI Lab, IBM Research. |
| Pseudocode | Yes | Algorithm 1 Heteroskedastic neural network with sufficiency-based regularizer. Algorithm 2 Residual-based neural network with calibration-based regularizer. |
| Open Source Code | Yes | 1The source code is available at github.com/Abhin02/ fair-selective-regression. |
| Open Datasets | Yes | We test our algorithms on Insurance and Crime datasets, and provide an application of our method in Causal Inference via IHDP dataset. The Insurance dataset (Lantz, 2019). Communities and Crime. The Crime dataset (Redmond & Baveja, 2002). The IHDP dataset (Hill, 2011). |
| Dataset Splits | No | We evaluate and report the empirical findings on a held-out test set with a train-test split ratio of 0.8/0.2. The paper specifies a train-test split, but does not explicitly mention a separate validation set split or how it was used for hyperparameter tuning. Therefore, it does not provide specific details for a train/validation/test split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments, such as GPU/CPU models, memory, or specific cloud computing instances. It only mentions 'neural networks' generally. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' and 'neural networks' but does not specify software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow versions, or specific library versions). |
| Experiment Setup | Yes | In all of our experiments, we use two-layer neural networks, and train our model only once on a fixed training set. For all hidden layers, we use the selu activation function. For the output layer, we use a non-linear activation function only for the variance-prediction network associated with Algorithm 2 to ensure that the predictions of variance are non-negative. In particular, we use the soft-plus activation function for the variance-prediction network associated with Algorithm 2. We train all our neural networks with the Adam optimizer, a batch size of 128, and over 40 epochs. Further, we use a step learning rate scheduler with an initial learning rate of 5 7 10 3 and decay it by a factor of half after every two epochs. Finally, as described in Section 6, we set the regularizer 0 = 1 throughout. |