FredericW / vldb_2024_submission

The full technical report is here

Implementation of A3M Mechanism

Dependencies

1. A3M needs the optimization module of scipy.
2. To run real-world data experiments (including multidimensional data), additional packages of pandas and folktable are needed.
3. Code for preprocessing the real-world data is in real_dataset.ipynb (you can also try multidimension.ipynb to play with the ACSData).

Synthetic data experiments

python synthetic_exp_dis.py

This script shows the performance of A3M, in comparison with several other baselines (run the script *_sw.py for SW) over a range of choices of epsilon. The data type and various hyperparameters (e.g., data size, bin size, split ratio) can be specified in the input arguments. For example, for 10000 Gaussian data points, run

python synthetic_exp_dis.py --data_type=GAUSSIAN --n=10000 --bin_size=0.5 --axRatio=4 --s=0.1

with bin_size (namely, N) set to 0.5; axRatio set to 4; and split ratio set to 0.1.

Real-world data experiments

python real_income.py

python real_retirement.py

python real_green_taxi.py

python multidimension.py

These scripts show the performance of A3M on multi-dimensional data for multiple real-world datasets, in comparison with several other baselines (run the script *_sw.py for SW).

Hyperparameter studies of A3M

Run AAA with different values of bin size (or equivalently, number of bins).

python vary_N_synthetic.py

Run AAA with different values of aX ratio.

python vary_ax_synthetic.py

Run AAA with different values of split portion.

python vary_split_synthetic.py