RELand: Risk Estimation of Landmines via Interpretable Invariant Risk Minimization

Mateo Dulce Rubio*, Siqi Zeng*, Qi Wang, Didier Alvarado, Francisco Moreno, Hoda Heidari, and Fei Fang

This repository contains all the instructions to replicate the results of the paper "RELand: Risk Estimation of Landmines via Interpretable Invariant Risk Minimization"

Requirements

Required packages:

pandas
numpy
sklearn
pytorch
lightgbm
pytorch_tabnet

Dataset

The dataset used for experiments in paper is in /processed_dataset folder.

Validation

We provide three validation methods in our paper in /train_val_stream folder.

• blockCV: blockCV method
• bolivar: blockV method
• transfer: transferCV method

Running Experiments

We provide an example in bash run_reland.sh. Some configurations include:

• --timestamp an unique experiment string
• --municipio use which validation method, either blockCV, bolivar or ``
• --subset use which subset of features, single (distance to historical landine), geo (geospatial features) or full (all 70 features)
• --model which model to use, TabCmpt is the RELand model, other options can be MLP, TabNet, LR, RF, SVM, LGBM
• --objective using irm, erm, or pnorm
• --n_step number of decision blocks
• --warm_start directory with checkpoints

Your final result for this run will be stored under /experiments/<timestamp> that contains

• all current .py files in the root directory
• a <municipality>.pth model for each municipality
• an <municipality>.png image for each municipality that visualizes the ground truth and prediction
• config.json with current configuration (hyper)parameters
• metrics.json that contains all 4 metrics
• predicted_proba.csv that combines validation prediction for all municipalities
• feature_importance.csv with global feature importance if applicable. Black-box models generate -1 values for all features.

Acknowledgment

We borrow and edit packages including ood-bench, scikit-learn/tree, pytorch-tabnet.