This is the code for reproducing the paper result. We propose a new method to gain feature importance from deep neural network.
Requirements:
python mlp_predict.py --dataset InteractionSimulation --rank_func nn_rank:0.1 nn_rank:0.5 nn_rank:1 nn_rank:0.05 marginal_rank rf_rank zero_rank shuffle_rank random_rank dfs_rank enet_rank lasso_rank
python mlp_predict.py --dataset NoInteractionSimulation --rank_func nn_rank:0.1 nn_rank:0.5 nn_rank:1 nn_rank:0.05 marginal_rank rf_rank zero_rank shuffle_rank random_rank dfs_rank enet_rank lasso_rank
* Support2 with zeroing out the feature
```bash
python mlp_predict.py --dataset support2 --identifier 0111 --rank_func all_rank --test_func nn_test_zero
python mlp_predict.py --dataset MiniBooNE --identifier 0111 \
--rank_func nn_rank:0.1 nn_rank:0.01 marginal_rank rf_rank zero_rank shuffle_rank \
random_rank dfs_rank enet_rank lasso_rank --test_func nn_test_zero
python mlp_predict.py --dataset YearMSD --identifier 0111 \
--rank_func nn_rank:1 marginal_rank rf_rank zero_rank shuffle_rank \
random_rank dfs_rank enet_rank lasso_rank nn_rank:0.1 --test_func nn_test_retrain
See notebooks for the further analysis and reproducing figures
Further questions?
CC 4.0 Attribution-NonCommercial International
The software is for educaitonal and academic research purpose only.
@article{chang2017dropout,
title={Dropout Feature Ranking for Deep Learning Models},
author={Chang, Chun-Hao and Rampasek, Ladislav and Goldenberg, Anna},
journal={arXiv preprint arXiv:1712.08645},
year={2017}
}