Distributionally Robust Performative Prediction
Authors: Songkai Xue, Yuekai Sun
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results demonstrate that DRPO offers potential advantages over traditional PO approach when the distribution map is misspecified at either microor macro-level. 5 Experiments We revisit Examples 2.1, 2.2, and 2.3 and compare the PO and the DRPO empirically. |
| Researcher Affiliation | Academia | Songkai Xue Department of Statistics University of Michigan sxue@umich.edu Yuekai Sun Department of Statistics University of Michigan yuekai@umich.edu |
| Pseudocode | Yes | Algorithm 1 DR Performative Risk Minimization and Algorithm 2 Tilted Performative Risk Minimization |
| Open Source Code | No | The data is modified from public dataset [16] or generated sythetically (see details in Section 5 and Appendix G) and the code is available upon request. |
| Open Datasets | Yes | For the base distribution D(0), we use a class-balanced subset of a Kaggle credit scoring dataset ([16], CC-BY 4.0). |
| Dataset Splits | No | The paper mentions generating training samples but does not specify explicit training, validation, or test dataset splits (e.g., percentages or counts) nor does it reference predefined splits with citations for reproducibility. |
| Hardware Specification | Yes | each replicate of the experiments is run on a local machine with an Intel Xeon Gold 6154 CPU and 32GB of RAM in less than an hour execution time. |
| Software Dependencies | No | The paper describes the model and loss function used but does not provide specific version numbers for software libraries or dependencies (e.g., Python, PyTorch, TensorFlow, scikit-learn versions). |
| Experiment Setup | Yes | For the loss function ℓ(x, y; θ), we adopt the cross-entropy loss with L2-regularization, that is, ℓ(x, y; θ) = y log hθ(x) (1 y) log(1 hθ(x)) + λ 2 θ 2 2 where hθ(x) = 1 + exp{ θ x} 1 and λ = 0.001. The data preprocessing procedure for the credit dataset [16] follows the procedure documented in [29]. After that procedure, the base distribution Dtrue(0) has n = 14878 data points with equal probability mass. We treat the distribution map associated with Dtrue(0) as the underlying test distribution map which is unknown to us. We generate n IID samples from Dtrue(0) to get a training base distribution b D(0), that is, b D(0) Dtrue(0) n, and then we have a training nominal distribution map induced by b D(0). |