ArXiv | BibTex

This is the code for reproducing the paper result. We propose a new method to gain feature importance from deep neural network.

0. Requirements:

   • Install python3
   • Install package pandas, seaborn, glmnet_python
   • Install pytorch 0.3 version (It does not support 0.4)
   • If you want to replicate the results of real datasets, download the datasets and put it under folder exp/DFRdatasets/data/.
     • Support2
     • MiniBooNE
     • Online News Popularity
     • Year Prediction MSD

1. Running examples:

   • Running simulation

   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

   python mlp_predict.py --dataset support2 --identifier 0111 --rank_func all_rank --test_func nn_test_zero

   • MiniBooNE

   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

   • Year MSD

   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

2. Further questions?

   • Post in the issue or send to my email [email protected]

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}
}