Code to reproduce the results from the following paper:

Ignatiadis, N., & Wager, S. (2019). Covariate-Powered Empirical Bayes Estimation. To appear in Advances in Neural Information Processing Systems 32 (NeurIPS 2019). arXiv:1906.01611.

The method itself has been implemented in the EBayes.jl package.

All computations were executed on Julia 1.2. The Manifest.toml file provides a specification of all dependencies. The exact package dependencies may thus be reproduced by starting the Julia REPL in this folder, pressing ] and then typing instantiate ..

There are three main results reproduced here:

• Synthetic experiments (simulations), i.e. we reproduce Figure 3 from the manuscript. Running julia simulations.jl runs all the simulations, evaluates the methods and stores the mean squared errors (± standard errors) in the synthetic_benchmark_files folder. For convenience we also provide precomputed results. The panels of Figure 2 may be generated afterwards by running julia synthetic_benchmark_plots.jl.
• MovieLens analysis, i.e. we reproduce Figure 2 from the manuscript. The corresponding code is available in the Jupyter notebook EBCrossFit_Movielens.ipynb. The preprocessing of the dataset occurs through the EBayesDatasets.jl package.
• Crimes and communities unnormalized analysis, i.e. we reproduce Table 1 from the manuscript. The code is available in the Jupyter notebook EBCrossFit_Crime_And_Communities.ipynb.