Approximate Selection with Guarantees using Proxies

This is the official project page for "Approximate Selection with Guarantees using Proxies."

Please read our paper for full technical details.

Requirements

You will need the following installed:

• python 3.x
• pandas
• numpy
• feather-format

You will also need to install the supg package. You can install by doing pip install supg, but we recommend using the latest version of the repository by running the following command in the root directory:

pip install -e .

Reproducing experiments

In supg/experiments, run experiments.py. This will reproduce many of the end-to-end numbers in the paper.

To run your own experiments, create a CSV with the following columns: id (0-indexed and sequential), label (True or False), and proxy_score (float between 0 and 1). You can run the precision or recall target experiments as follows:

from supg import run_rt, run_pt

run_pt(csv_fname, budget, rt)  # Precision target
run_rt(csv_fname, budget, rt)  # Recall target