hawks 0.2.0

A package for generating synthetic clusters, with parameters to customize different aspects of the complexity of the cluster structure

HAWKS is a tool for generating controllably difficult synthetic data, used primarily for clustering.

This repo is associated with the following paper:

1. Shand, C., Allmendinger, R., Handl, J., Webb, A., & Keane, J. (2019, July). Evolving controllably difficult datasets for clustering. In Proceedings of the Genetic and Evolutionary Computation Conference (pp. 463-471). https://doi.org/10.1145/3321707.3321761 (Nominated for best paper on the evolutionary machine learning track at GECCO’19)

The academic/technical details can be found there. What follows here is a practical guide to using HAWKS to generate synthetic data.

If you use HAWKS to generate data that forms part of a paper, please cite the paper above and link to this repo.

Installation

Installation is available through pip by:

pip install hawks

or by cloning this repo (and installing locally using pip install .). HAWKS was written for Python 3.6+. Other dependencies are specified in the setup.py file.

Running HAWKS

The parameters of hawks are configured via a config file system. Details of the parameters are found in the documentation. For any parameters that are not specified, default values will be used (as defined in hawks/defaults.json).

The example below illustrates how to run hawks. Either a dictionary or a path to a JSON config can be provided to override any of the default values. Further examples can be found here.

"""Single, simple HAWKS run, with KMeans applied to the best dataset
"""
import numpy as np
from sklearn.cluster import KMeans
from sklearn.metrics import adjusted_rand_score
import hawks

# Set the magic seed number
SEED_NUM = 42
# Set the seed number in the config
config = {
    "hawks": {
        "folder_name": "simple_example",
        "seed_num": SEED_NUM
    },
    "dataset": {
        "num_clusters": 5
    },
    "objectives": {
        "silhouette": {
            "target": 0.9
        }
    }
}
# Any missing parameters will take from hawks/defaults.json
generator = hawks.create_generator(config)
# Run the generator
generator.run()
# Let's plot the best individual found
generator.plot_best_indivs(show=True)
# Get the best dataset found and it's labels
datasets, label_sets = generator.get_best_dataset()
# Stored as a list for multiple runs
data, labels = datasets[0], label_sets[0]
# Run KMeans on the data
km = KMeans(
    n_clusters=len(np.unique(labels)), random_state=SEED_NUM
).fit(data)
# Plot the output of KMeans
hawks.plotting.scatter_prediction(data, km.labels_)
# Get the Adjusted Rand Index for KMeans on the data
ari = adjusted_rand_score(labels, km.labels_)
print(f"ARI: {ari}")

Documentation

For further information about how to use HAWKS, including examples, please see the documentation.

Issues

As this work is still in development, plain sailing is not guaranteed. If you encounter an issue, first ensure that hawks is running as intended by navigating to the tests directory, and running python tests.py. If any test fails, please add details of this alongside your original problem to an issue on the GitHub repo.

Contributing

At present, this is primarily academic work, so future developments will be released here after they have been published. If you have any suggestions or simple feature requests for HAWKS as a tool to use, please raise that on the GitHub repo.

I have various directions for HAWKS at present, and can only work on a subset of them, and so involvement with more people would be great. If you would like to extend this work or collaborate, please contact me.