Statistically Valid Variable Importance Assessment through Conditional Permutations

Authors: Ahmad CHAMMA, Denis A. Engemann, Bertrand Thirion

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments & Results, We conduct extensive benchmarks on synthetic and heterogeneous multimodal real-world biomedical data tapping into different correlation levels and data-generating scenarios for both classification and regression (section 5).
Researcher Affiliation Collaboration Ahmad Chamma Inria, Universite Paris Saclay, CEA ahmad.chamma@inria.fr Denis A. Engemann Roche Pharma Research and Early Development, Neuroscience and Rare Diseases, Roche Innovation Center Basel, F. Hoffmann La Roche Ltd., Basel, Switzerland denis.engemann@roche.com Bertrand Thirion Inria, Universite Paris Saclay, CEA bertrand.thirion@inria.fr
Pseudocode Yes Algorithm 1 Conditional sampling step: The algorithm implements the conditional sampling step in place of the permutation approach when computing the p-value of variable xj
Open Source Code Yes We propose a reusable library for simulation experiments and real-world applications of our method on a public Git Hub repo https://github.com/achamma723/Variable_ Importance.
Open Datasets Yes A recent real-world data analysis of the UK Biobank dataset reported successful machine learning analysis of individual characteristics. The UK Biobank project (UKBB) curates phenotypic and imaging data from a prospective cohort of volunteers drawn from the general population of the UK [Constantinescu et al., 2022]. Age prediction from brain activity (MEG) in Cam-CAN dataset Following the work of Engemann et al. [2020], we have applied CPI-DNN to the problem of age prediction from brain activity in different frequencies recorded with magnetoencephalography (MEG) in the Cam-CAN dataset.
Dataset Splits Yes Throughout the paper, we rely on an i.i.d. sampling train / test partition scheme where the n samples are divided into ntrain training and ntest test samples and our implementation involves a 2-fold internal validation (the training set of further split to get validation set for hyperparameter tuning).
Hardware Specification No The paper mentions 'per core on 100 cores' when discussing computation time, but it does not specify the type or model of CPU/GPU or any other detailed hardware specifications used for the experiments.
Software Dependencies No The paper mentions general software components like deep neural networks and random forests, but it does not specify any software libraries with version numbers (e.g., Python, PyTorch, scikit-learn versions) required to replicate the experiments.
Experiment Setup No The paper states that 'hyperparameter tuning' and '2-fold internal validation' were used for models like Random Forests (e.g., 'the max depth of the Random Forest is chosen based on the performance with 2-fold cross validation'), but it does not explicitly provide the specific values for these hyperparameters or other system-level training settings in the main text.