Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Mixed-curvature decision trees and random forests
Authors: Philippe Chlenski, Quentin Chu, Raiyan R. Khan, Kaizhu Du, Antonio Khalil Moretti, Itsik Pe’Er
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In benchmarks on a diverse suite of 57 classification, regression, and link prediction tasks, our product RFs ranked first on 29 tasks and came in the top 2 for 41. This highlights the value of product RFs as straightforward yet powerful new tools for data analysis in product manifolds. |
| Researcher Affiliation | Academia | 1Columbia University 2Barnard College 3Spelman College. Correspondence to: Philippe Chlenski <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Product Space Decision Tree |
| Open Source Code | Yes | Code for our method is available at https://github.com/pchlenski/manify. |
| Open Datasets | Yes | Appendix J. Datasets availability. This table lists all of the datasets used in this paper, with download links and citations. For example, Cite Seer (Giles et al., 1998) from Network Repository, MNIST (Lecun et al., 1998) from Hugging Face, Traffic (Fedesoriano, 2020) from Kaggle. |
| Dataset Splits | Yes | We apply an identical 80:20 train-test split to all of our data, train our models on the training set, and evaluate performance on the test set. |
| Hardware Specification | No | The paper mentions running experiments and evaluating models but does not provide specific details on the hardware used (e.g., specific GPU or CPU models, memory sizes). |
| Software Dependencies | No | The paper mentions several software packages and libraries such as Scikit-Learn, Manify, Geoopt, PyTorch (implicitly for Geoopt), NetworkX, and Matplotlib. However, it does not provide specific version numbers for any of these dependencies, which is required for a reproducible description. |
| Experiment Setup | Yes | Specifically, we set the following hyperparameters for both DTs and RFs: max depth = 5 min samples split = 2 min samples leaf = 1 min impurity decrease = 0.0. For RFs, we also set the following hyperparameters: n estimators = 12 max features = "sqrt" bootstrap = True (subsamples the training data) max samples = None. For all neural networks, we used a learning rate of .0001 and trained for 4,000 epochs. Both optimizers use the hyperparameters β1 = 0.9 and β2 = 0.999. Batch size: 4,096 Number of samples per point: 64 β (weight for KL-divergence in VAE loss): 1. |