hang-wu/l2l-scrm

Sequential Counterfactual Risk Minimization

This repository contains the code for reproducing experiments of the paper Sequential Counterfactual Risk Minimization submitted at ICML.

Prerequisites

Software:

pip install -U -r requirements.txt

Download discrete datasets from the LibSVM website:

BASEURL="https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multilabel"
wget $BASEURL/yeast_test.svm.bz2
wget $BASEURL/yeast_train.svm.bz2
wget $BASEURL/scene_test.bz2
wget $BASEURL/scene_train.bz2
wget $BASEURL/tmc2007_test.svm.bz2
wget $BASEURL/tmc2007_train.svm.bz2

bunzip2 *bz2

Instructions

Discrete Experiments

bash generate_figures_and_tables.sh

Continuous Experiments

Go into the continuoussub folder and run the following.

Figure 1: run the Jupyter notebook in continuous/

gaussian_example.ipynb

Figure 2 SCRM vs CRM, continuous datasets

python continuous/scrm_vs_crm.py

Table 1 SCRM vs CRM, example 3.1

python continuous/gaussian_example.py

Table 3 SCRM vs baselines, continuous datasets

python continuous/scrm_vs_baselines_counterfactual.py
python continuous/scrm_vs_baselines_sbpe.py
python continuous/scrm_vs_baselines_bkucb.py
python continuous/scrm_vs_rl.py

Figure 3,10: run the Jupyter notebook

compare_estimators.ipynb

Figure 4: run the Jupyter notebook in continuous/

distance_scrm_gaussian_example.ipynb

Batch bandits Baselines

Discrete datasets

python batch_bandits_experiments.py

Continuous datasets

python continuous/scrm_vs_baselines_sbpe.py
python continuous/scrm_vs_baselines_bkucb.py

RL Baselines

Discrete datasets

python compare_rl_scrm.py scene,yeast,tmc2007

Continuous datasets

python continuous/scrm_vs_rl.py

Figure 8: run the Jupyter notebook in continuous/

holderian_error_bound_assumption.ipynb