Logistic Regression Plus-Minus

This repository contains the codes for David Poole, Ali Mohammad Mehr, Wan Shing Martin Wang ,"Conditioning on “and nothing else”: Simple Models of Missing Data betweenNaive Bayes and Logistic Regression," ICML 2020 Workshop Artemiss Submission.

Running the code

1. pip install -r requirements.txt

2. ./run_tests.sh 1 200 10

Explanations for above commands

pip install -r requirements.txt will install Python dependencies for the code. The main dependencies include sklearn and pymc3.

Running ./run_tests.sh 1 200 10 will run the tests and produce the output graphs. As you can see, this script accepts 3 arguments as follows:

• start_id: The id of the first test. Suggested value: 1

• end_id: The id of the last test. Suggested value: 200

• Number of concurrent tests running. Suggested value: 10 (if your system has at least 10 CPU threads)

The script will run $$(end_id-start_id+1)$$ tests. Each test is composed of creating 3 randomly generated datasets - namely a C dataset, a D4 dataset, and a DLR dataset. On each generated dataset, the code traines the four different models mentioned in our paper - namely models LR$$\pm$$, Naive Bayes, model (c), and model (d). For each generated dataset, the logloss of each model for that dataset is stored in an individual json file in results/ directory.

In the end, the script will run python Read-test-results-and-plot-graphs.py to make the graphs shown in paper alongs with printing average logloss comparisons.

Using the suggested argument values for ./run_tests.sh will result in running 200 tests which are run simultaneously on 10 CPU threads. Note that it will take more than 6 hours for the code to run 200 tests simultaneously on 10 CPU threads.