Robust Loss Functions for Training Decision Trees with Noisy Labels
Authors: Jonathan Wilton, Nan Ye
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Lastly, our experiments on multiple datasets and noise settings validate our theoretical insight and the effectiveness of our adaptive negative exponential loss. |
| Researcher Affiliation | Academia | School of Mathematics and Physics, The University of Queensland |
| Pseudocode | Yes | See Algorithm 1 in Appendix J for details. |
| Open Source Code | Yes | Our source code is available at https://github.com/jonathanwilton/Robust Decision Trees. |
| Open Datasets | Yes | We consider some commonly used datasets from UCI including Covertype (Blackard 1998), 20News (Lang 1995), Mushrooms (Audubon Society Field Guide 1987), as well as the MNIST digits (Le Cun et al. 1998), CIFAR-10 (Krizhevsky 2009) and UNSW-NB15 (Moustafa and Slay 2015). |
| Dataset Splits | Yes | The NE impurity hyperparameter λ {0, 0.25, 0.5, 0.75, 1} was tuned by training on 80% of the noisy training set and validation on the other 20%. |
| Hardware Specification | Yes | Experiments were performed using Python on a computer cluster with Intel Xeon E5 Family CPUs @ 2.20GHz and 192GB memory running Cent OS. |
| Software Dependencies | No | The paper mentions 'Python' and 'scikit-learn' but does not provide specific version numbers for these software components, which are required for a reproducible description. |
| Experiment Setup | Yes | Following common practice (Pedregosa et al. 2011), we set no restriction on maximum depth or number of leaf nodes, minimum one sample per leaf node, 100 trees in each RF, each using only d randomly chosen features, and predictions with RF made by majority average label distribution over all trees. The NE impurity hyperparameter λ {0, 0.25, 0.5, 0.75, 1} was tuned by training on 80% of the noisy training set and validation on the other 20%. GCE q = 0.7 as recommended for NNs (Zhang and Sabuncu 2018), Credal-C4.5 s = 1 as recommended in (Mantas and Abellan 2014). |