Experiments
This section documents how to reproduce the experiments run with the HGP library on the datasets used in our evaluation.
Each page covers, where applicable:
- how to obtain the raw data
- how to preprocess it into the HDF format consumed by the library
- how to run a benchmark with
scripts/run_benchmark.py
Datasets
- PMLB: a suite of binary classification benchmarks, used to evaluate HGP across many small to medium datasets.
- PaySim: synthetic mobile-money transactions for fraud detection, large and strongly imbalanced.
- AEAC: Amazon employee access requests, where readable rules double as access-control policies.
Conventions
All scripts use a consistent --data_path argument to point at the dataset.
Preprocessing scripts accept --data_path as the folder that holds the raw input and receives the generated .hdf files.
scripts/run_benchmark.py and scripts/optuna_hypertuning.py take --data_path as the path to a single preprocessed .hdf file.