Conformal Off-Policy Prediction

This repository is the official implementation of of Conformal Off-policy Prediction. It includes the toy example in Section 3.2 and the synthetic data analysis in Section 5.

Requirements:

The experiments is run in R 4.0.3. It requires R packages "grf", "gbm", "bartMachine","randomForest","glmnet","ggplot2"

Training:

1. Toy example is implemented in the folder Figure1: DirectM.R: the function for direct method, we can choose true/false model, deterministic/random policy SS_low_DirectM.R: run 100 repetitions for Direct method

   SamplingM.R: the function for subsampling method,we can choose true/false model, deterministic/random policy SS_low_SamplingM.R: run 100 repetitions for Subsampling method

   AOPM.R: the function for COPP method,we can choose true/false model, deterministic/random policy SS_low_AOPM.R: run 100 repetitions for COPP method

   conformal_utils.R: functions for conformal prediction genY.R: data generation

   All training setups in SS_low_DirectM.R, SS_low_SamplingM.R, SS_low_AOPM.R are the same with the paper.

2. Synthetic data analysis is implemented in the folder Figure2.

   conformal_learners.R: prediction models such as quantile random forest
   conformal_utils.R: functions for weighted conformal prediction, construct CI conformalCI.R: the main function COPP for single-stage decision making, Algorithm 1 in main paper conformalCI_TS.R: the main function COPP for multi-stage decision making, Alogorithm 1 in supp genY.R: data generation propensityscore.R: prediction models for propensity score such as logistic regression quantilereg.R: convert a valid outfun string to the function samplesplit.R: split samples randomly as training and calibration sets

   SS_high_BootDweight.R: high dim,single-stage, implement 100 repetitions of the COPP_IS_MS method(B=100, method="BootDweight"), COPP_IS(B=1, method="Dweight") SS_high_BootSweight.R: high dim,single-stage, implement 100 repetitions of the COPP_MS method(B=100, method="BootSweight"), COPP(B=1, method="Sweight") SS_high_naive.R: high dim,single-stage, implement 100 repetitions of the Subsampling-based method

   SS_low_BootDweight.R: low dim,single-stage, implement 100 repetitions of COPP_IS_MS method(B=100, method="BootDweight"), COPP_IS(B=1, method="Dweight") SS_low_BootSweight.R: low dim,single-stage, implement 100 repetitions of COPP_MS method(B=100, method="BootSweight"), COPP(B=1, method="Sweight") SS_low_naive.R: low dim,single-stage, implement 100 repetitions of Subsampling-based method SS_low_IPWE.R: low dim,single-stage, implement 100 repetitions of IS and DR method

   TS_XXX_XXX.R is the same as SS but for two-stage setting

Results：

All results have been described in our paper. The boxplots in the paper are summarized by the 100 repetitions in the summary.R in Figure1 and summary_high.R, summary_low.R in Figure2.

Contact: yyzhang@fem.ecnu.edu.cn