Causal Strategic Classification: A Tale of Two Shifts
Authors: Guy Horowitz, Nir Rosenfeld
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on synthetic and semi-synthetic data demonstrate the utility of our approach. ... Finally, we conduct a series of experiments that empirically validate our approach. |
| Researcher Affiliation | Academia | 1Technion Israel Institute of Technology, Haifa, Israel. Correspondence to: Nir Rosenfeld <nirr@cs.technion.ac.il>. |
| Pseudocode | Yes | Pseudocode for our entire procedure is given in Algorithm (1). |
| Open Source Code | Yes | All code is made publicly available and can be found at: https://github.com/guyhorowitz/CSC. |
| Open Datasets | Yes | We now turn to experiments based on real data using two public datasets: (i) spam, used originally in Hardt et al. (2016), and (ii) card fraud, used in Levanon & Rosenfeld (2021). ... The data is publicly available at https://www.kaggle.com/datasets/mlg-ulb/ creditcardfraud. |
| Dataset Splits | Yes | Next, we split the data roughly 60-10-30 into train, validation, and test sets. ... We sampled 5500 balanced samples for the experiment and set 3000 of them (~54%) as training data, 500 (~9%) as validation data, and 2000 (~36%) as test data. |
| Hardware Specification | No | The paper provides details on data processing, feature partition, labeling functions, and training parameters (e.g., learning rates, batch size, epochs), but it does not specify the hardware (e.g., GPU/CPU models) used for running the experiments. |
| Software Dependencies | No | The paper mentions using KDE with a Gaussian kernel and implies Python for the open-sourced code, but it does not provide specific version numbers for software dependencies or libraries (e.g., PyTorch, scikit-learn). |
| Experiment Setup | Yes | For both card fraud and spam experiments, we used the following parameters, which we choose manually: ... 2. f, h learning-rate = 0.01 3. batch size = 64 4. epochs = 100 5. an early stopping mechanism when there are 7 consecutive epochs without accuracy improvement on the validation set 6. sigmoid temperature τ = 4 7. exploration regularization coefficients: in CSERMλ = 0.1 we used λ0 = 0.1 which decays in each round with factor of 0.4. in CSERMλ = 1 we used λ0 = 1 which decays in each round with factor of 0.4. |