Using Noise to Infer Aspects of Simplicity Without Learning
Authors: Zachery Boner, Harry Chen, Lesia Semenova, Ronald Parr, Cynthia Rudin
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We confirm our results empirically and provide practical guidance for the datasets from domains of criminal justice and lending, where we expect outcome noise due to the random nature of the data generation process. (Section 1, page 2) and We now present experimental results supporting the results in Section 4 for uniform label noise and 0-1 loss and Section 6 for additive attribute noise and exponential loss. (Section 7, page 5) |
| Researcher Affiliation | Collaboration | Zachery Boner1 Harry Chen1 Lesia Semenova2 Ronald Parr1 Cynthia Rudin1 1Department of Computer Science, Duke University, 2Microsoft Research (Page 1) |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code and data may be found in the supplemental materials. (NeurIPS Paper Checklist, Open access to data and code, Justification, page 14) |
| Open Datasets | Yes | In Table 2, we provide the description of the datasets used in this paper and pre-processing steps. (Section I, page 10). Table 3: Licensing and Data Source Information for all Datasets (Table 3, page 11) showing specific citations like Amsterdam (Recidivism) DANS Tollenaar and Heijden [2013] and NIJ Recidivism Challenge Publicly Available NIJ [2011]. |
| Dataset Splits | Yes | Before injecting any noise, the optimal value λ of the regularization parameter for the GOSDT algorithm was chosen using 5-fold cross validation on the training set. (Section I.2, page 11) |
| Hardware Specification | No | We performed experiments on Duke University s Computer Science Department cluster. We requested 200GB of shared total memory, and one compute core per dataset (23 datasets) so that we could run the experiments for each dataset in parallel. (Section I.5, page 13) |
| Software Dependencies | No | We used the GOSDT-guesses algorithm by Mc Tavish et al. [2022] to optimize sparse decision trees over varying amounts of label noise (between 0.0 and 0.3). (Section 7.1, page 5). We optimize for 1000 epochs with an initial learning rate of 0.1 with the ADAM optimizer [Kingma and Ba, 2015]. (Section 7.2, page 6). No specific version numbers for these software components are provided. |
| Experiment Setup | Yes | Throughout the experiment, we allowed a maximum depth of 5... the optimal value λ of the regularization parameter for the GOSDT algorithm was chosen using 5-fold cross validation... for 51 parameters ρ for label noise linearly spaced between 0 and 0.3... Sample 250 i.i.d. draws of random label noise... We optimize for 1000 epochs with an initial learning rate of 0.1 with the ADAM optimizer [Kingma and Ba, 2015]. We furthermore decrease the learning rate by a factor of 0.3 if the exponential loss plateaus for over 50 epochs. (Section I.2, page 11-12). multiplicative Rashomon parameter of 0.05 and a regularization parameter of 0.02 throughout these experiments. (Section I.1, page 11). |