Multicalibration as Boosting for Regression
Authors: Ira Globus-Harris, Declan Harrison, Michael Kearns, Aaron Roth, Jessica Sorrell
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally we investigate the empirical performance of our algorithm experimentally using an open source implementation that we make available on Git Hub. We give a fast, parallelizable implementation of our algorithm and in Section 7 demonstrate its convergence to Bayes optimality on two-dimensional datasets useful for visualization, as well as evaluate the accuracy and calibration guarantees of our algorithm on real Census derived data using the Folktables package (Ding et al., 2021). |
| Researcher Affiliation | Academia | Department of Computer and Information Sciences, University of Pennsylvania, Philadelphia PA, USA. |
| Pseudocode | Yes | Algorithm 1 LSBoost(f, α, AH, D, B) |
| Open Source Code | Yes | Finally we investigate the empirical performance of our algorithm experimentally using an open source implementation that we make available on Git Hub 2. Our code repository can be found at https://github. com/Declancharrison/Level-Set-Boosting |
| Open Datasets | Yes | We evaluate the empirical performance of Algorithm 1 on US Census data compiled using the Python folktables package (Ding et al., 2021). |
| Dataset Splits | No | On an 80/20% train-test split with 500,000 total samples, we compare the performance of Algorithm 1 with Gradient Boosting with two performance metrics: mean squared error (MSE), and mean squared calibration error (MSCE). The paper specifies a train-test split but does not explicitly mention a separate validation split or strategy. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types, memory specifications). It only mentions general terms like "parallelizable implementation" and discusses training times. |
| Software Dependencies | No | The paper mentions using "scikit-learn" and a "Python implementation" but does not provide specific version numbers for these or any other key software dependencies (e.g., Python version, scikit-learn version, PyTorch/TensorFlow versions, CUDA versions, etc.). |
| Experiment Setup | Yes | In Figure 4, we show an example of Algorithm 1 learning C0 using a discretization of five-hundred level sets and a weak learner hypothesis class of depth one decision trees. On an 80/20% train-test split with 500,000 total samples. |