Strategic Classification Made Practical
Authors: Sagi Levanon, Nir Rosenfeld
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A series of experiments demonstrates the effectiveness of our approach on various learning settings. We conduct a series of experiments demonstrating the effectiveness of our approach. With respect to the above points, each of our experiments is designed to study a different practical aspect of learning. The experiments cover a range of learning environments using real and synthetic data. |
| Researcher Affiliation | Academia | Department of Computer Science, Technion Israel Institute of Technology. Correspondence to: Nir Rosenfeld <nirr@cs.technion.ac.il>. |
| Pseudocode | No | The paper describes procedures and algorithms in prose but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | Towards this end, we make publicly available a code repository with a flexible implementation of our approach, designed to support a wide range of strategic learning settings. Code can be found at https://github.com/SagiLevanon1/scmp. |
| Open Datasets | Yes | Across all experiments we use four real datasets: (i) spam7, which includes features describing users of a large social network, some of which are spammers (used originally in Hardt et al. (2016)); (ii) credit8, which includes features describing credit card spending patterns, and labels indicating default on payment (we use the version from Ustun et al. (2019)); (iii) fraud9, which includes credit card transactions that are either genuine or fraudulent (Dal Pozzolo et al., 2015); and (iv) financial distress10, which includes time-series data describing businesses over time along with labels indicating their level of financial distress and whether they have gone bankrupt. All datasets include features that describe users and relate to tasks in which users have incentive to obtain positive predictive outcomes. Some experiments use synthetic environments, described below. (Footnotes 7, 8, 9, 10 provide citations/links to the datasets). |
| Dataset Splits | Yes | All methods use a 60-20-20 data split, and results are averaged over multiple random splits. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments, such as GPU/CPU models, memory, or cloud instance types. |
| Software Dependencies | No | The paper mentions using 'Adam' for optimization, 'recursive neural networks (RNN)', 'convex optimization layers', and the 'differentiable convex solver of Agrawal et al. (2019a)' but does not provide specific version numbers for any software or libraries (e.g., Python, PyTorch, TensorFlow). |
| Experiment Setup | Yes | For optimizing our approach we use Adam with randomized batches and early stopping (most runs converged after at most 7 epochs). User responses were simulated using from Eq. (1) with τ = 1 for training and τ = 0.2 for evaluation. Tolerance for CCP convergence was set to 0.01 (most attempts converged after at most 5 iterations). |