Being Properly Improper
Authors: Tyler Sypherd, Richard Nock, Lalitha Sankar
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We study the twist-proper α-loss under a novel boosting algorithm, called PILBOOST, and provide formal and experimental results for this algorithm. Our overarching practical conclusion is that the twistproper α-loss outperforms the proper log-loss on several variants of twisted data. In Section 6, we implement PILBOOST with the approximate inverse canonical link of α-loss on several tabular datasets, each suffering from various twists (label, feature, and adversarial noise), and compare against Ada Boost (Freund & Schapire, 1997) and XGBoost (Chen & Guestrin, 2016). |
| Researcher Affiliation | Collaboration | 1School of Electrical, Computer and Energy Engineering, Arizona State University; 2Google Research. Correspondence to: Tyler Sypherd <tsypherd@asu.edu>. |
| Pseudocode | Yes | Algorithm 1 PILBOOST |
| Open Source Code | Yes | The code for all of our experiments (including the implementation of PILBOOST) can be found at the following github repository link: https://github.com/Sankar Lab/Being-Properly-Improper |
| Open Datasets | Yes | We provide experimental results on PILBOOST (for α {1.1, 2, 4}) and compare with Ada Boost (Freund & Schapire, 1997) and XGBoost (Chen & Guestrin, 2016) on four binary classification datasets, namely, cancer (Wolberg et al., 1995), xd6 (Buntine & Niblett, 1992), diabetes (Smith et al., 1988), and online shoppers intention (Sakar et al., 2019). |
| Dataset Splits | No | The paper mentions 'train/test split' and 'cross-validation' but does not explicitly describe a separate 'validation' dataset split for hyperparameter tuning. |
| Hardware Specification | Yes | Most of the experiments were performed over the course of a month on a 2015 Mac Book Pro with a 2.2 GHz Quad-Core Intel Core i7 processor and 16GB of memory. The Adaptive α experiments were performed on a computing cluster and each required about 30 minutes of compute time. |
| Software Dependencies | No | The paper mentions using decision trees and XGBoost but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | All algorithms across all experiments ran for 1000 iterations. For α = 1.1, 2, and 4, we set af = 7, 2, and 4, respectively. Hyperparameters of XGBoost were kept to default to maintain the fairest comparison between the three algorithms; for more of these experimental details, please refer to Appendix B.5. All experiments use regression decision trees (of varying depths 1-3) in order to align with XGBoost. |