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].
Interpretable Mesomorphic Networks for Tabular Data
Authors: Arlind Kadra, Sebastian Pineda Arango, Josif Grabocka
NeurIPS 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we demonstrate that our explainable deep networks have comparable performance to state-of-the-art classifiers on tabular data and outperform current existing methods that are explainable by design. |
| Researcher Affiliation | Academia | Arlind Kadra Department of Representation Learning University of Freiburg EMAIL Sebastian Pineda Arango Department of Representation Learning University of Freiburg EMAIL Josif Grabocka Department of Machine Learning University of Technology Nuremberg EMAIL |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We make our implementation publicly available4. (footnote 4: Source code at https://github.com/Arlind Kadra/IMN) |
| Open Datasets | Yes | We run our predictive accuracy experiments on the Auto ML benchmark that includes 35 diverse classification problems... For more details about the datasets included in our experiments, we point the reader to Appendix C. (Appendix C includes Table 6: Statistics regarding the Auto ML benchmark datasets, with Dataset IDs which can be accessed from Open ML) |
| Dataset Splits | Yes | All datasets have a train/validation set ratio of 10 to 1. |
| Hardware Specification | Yes | Lastly, the methods that offer GPU support are run on a single NVIDIA RTX2080Ti, while, the rest of the methods are run on an AMD EPYC 7502 32-core processor. |
| Software Dependencies | No | The paper mentions using PyTorch as the main library, scikit-learn, Optuna, and specific implementations for CatBoost, TabNet, and NAM. However, it does not provide specific version numbers for these software dependencies (e.g., 'PyTorch 1.x' or 'scikit-learn 0.y'). |
| Experiment Setup | Yes | For the default hyperparameters of our method, we use 2 residual blocks and 128 units per layer combined with the GELU activation (Hendrycks & Gimpel, 2016). When training our network, we use snapshot ensembling (Huang et al., 2017) combined with cosine annealing with restarts (Loshchilov & Hutter, 2019). We use a learning rate and weight decay value of 0.01, where, the learning rate is warmed up to 0.01 for the first 5 epochs, a dropout value of 0.25, and an L1 penalty of 0.1 on the weights. Our network is trained for 500 epochs with a batch size of 64. |