From Search to Sampling: Generative Models for Robust Algorithmic Recourse
Authors: Prateek Garg, Lokesh Nagalapatti, Sunita Sarawagi
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We then evaluate on three large real-life datasets that are popularly used in the recourse literature and compare our results with eight existing methods. We show that our method achieves (1) the best score combining validity, proximity and plausibility, (2) is more robust to changes in the cost magnitude compared to SOTA likelihood-based methods, and (3) more faithfully learns the conditional density of recourse than methods that separately learn unconditional data density. |
| Researcher Affiliation | Academia | Prateek Garg Lokesh Nagalapatti Sunita Sarawagi Indian Institute of Technology Bombay |
| Pseudocode | Yes | Algorithm 1 Training Recourse model Rθ Require: Training data D, lrn rate η, batch size b, epochs e, sampling parameter K, cost coefficient λ, pair validity γ, classifier h Ensure: Trained model Rθ with parameters θ |
| Open Source Code | Yes | Our code is available at: https://github.com/prateekgargx/genre. |
| Open Datasets | Yes | We experimented with benchmark datasets commonly used to evaluate recourse algorithms: Adult Income (Becker & Kohavi, 1996), FICO HELOC (FICO, 2018), and COMPAS (Angwin et al., 2016). |
| Dataset Splits | Yes | For each dataset, we train a Random Forest (RF) classifier to mimic the latent decisionmaking model, which assigns the gold labels... Adult Income 13 7 2 8,742 27,877 36624 12208 77.33 COMPAS 7 4 2 3,764 865 4629 1543 69.60 HELOC 21 0 0 3,548 3,855 7403 2468 74.23 |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models or memory amounts used for experiments. |
| Software Dependencies | No | We use a Transformer (Vaswani et al., 2017) from Py Torch (Paszke et al., 2019) with learned position embedding... We ensure that the RF classifier is calibrated by using the Calibrated Classifier CV API from sklearn (Pedregosa et al., 2011). The paper mentions software such as PyTorch and scikit-learn, but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | The classifier h(x) is an Artificial Neural Network (ANN) a Re LU-based model with three hidden layers of size 10 each, trained with a learning rate of 0.001 for 100 epochs using a batch size of 64... For training Rθ, we use a Transformer (Vaswani et al., 2017) from Py Torch (Paszke et al., 2019) with learned position embedding, embedding size 32, and 16 layers in each of encoder and decoder, and 8 heads. The number of bins in the last layer is 50. We choose the value of λ = 5.0 when sampling training pairs. During inference (Algorithm 2), we set the temperature for bin selection τ = 10.00 and σ = 0.00, generate 10 samples and choose the sample which gets highest probability from the classifier h(x). In Appendix D.2, we provide results over other values of τ and σ. We describe other relevant hyperparameters in Appendix C.2. All datasets uses learning rate = 1e-4 and batch size 16384 for Adult Income dataset else 2048. |