ma1VAR3 / DAPaMean-MD

Distribution Adaptive Private Mean Estimation - Multi Dimensional

DAPaMean

The paper on this research is live on arXiv: https://arxiv.org/abs/2401.15906

Docs

This repository contains code to run experiments for different setting available for the framework DAPaMean (single dimension).

The following files are crucial for running:

• config.json : Contains the configurations for the experiments (make changes to this file only)

• main.py : Contains the main pipeline, reads configurations from the config.json

• filter_data.py : Contains script to generate and store synthetic dataset in the gen_data directory

Steps:

1. Install all the requirements (requirements.txt)
2. Place the dataset in data/
3. Make necesarry changes in config.json
4. Run filter_data.py
5. Run main.py

NOTE: There are 3 possible scenarios (original data, synthetic data in which number of samples are scaled, synthetic data in which number of users are scaled). You need to run the filter_data.py only ONCE for EACH scenario. For a given scenario, to execute different experiments (6 possible experiments), simply change settings in the config.json and run main.py.

Possible experiments

concentration_algorithm	user_groupping
baseline2	wrap
baseline2	best_fit
coarse_mean	wrap
coarse_mean	best_fit
quantiles	wrap
quantiles	best_fit

The config.json file

The following parameters need to be taken care of before running the experiments:

• Synthetic : true/false (If true then synthetic data is used. Note that you should have generated the data using filter_data.py)

• Synthetic Factor : samples/users (Relevant only when "Synthetic" field is true. "samples" generates synthetic data in which number of samples are scaled. "users" generated synthetic data in which number of users are scaled.)

• Synthetic Factor : A number (The factor by which the number of samples are scaled. Should be kept the same for all the experiments for homogeneity)

• epsilons : A list of numbers (The epsilons to be tested. Must be the same for all experiments)

• num_experiments : A number (The number of times each expeirment is conducted. Myst be the same for all experiments)

• user_grouppung : wrap/best_fit (The grouping algo to be used)

• concentration_algorithm : baseline2/coarse_mean/quantiles (The estimation algorithm to be used)