A Python implementation of various metrics (primarily external) used for clustering evaluation. The documentation is available online here.

After creating an in-memory representation of a clustering or a partition, many common metrics can be calculated very cheaply. The efficiency of the computation depends primarily on the in-memory representation of clustering. Sparse representations are pefect for this purpose and allow us to calculate many metrics more efficiently than packages like Scikit-Learn.

At the moment, the package is not on PyPI. To install it, use pip like so:

pip install git+https://github.com/escherba/pymaptools#egg=pymaptools-0.2.31
pip install git+https://github.com/escherba/clustering-metrics#egg=clustering_metrics-0.0.2

Clusters can be represented in different ways. One way is to enumerate all items in the cluster with integer labels:

>>> ground_truth = [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2]
>>> predicted = [0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 2, 2, 1, 2, 2, 2]

Note that ground_truth and predicted must have the same length. We can then produce various metrics as follows:

>>> from clustering_metrics.metrics import ClusteringMetrics
>>> cm = ClusteringMetrics.from_labels(ground_truth, predicted)
>>> cm.adjusted_rand_index()
0.242914979757085

Another way to represent clusters is using partition-style encoding. Here, each clustering is represented as a set of partitions:

>>> ground_truth = [{1, 2, 3, 4}, {5, 6, 7, 8, 9, 10}, {11, 12, 13, 14, 15, 16}]
>>> predicted = [{1, 2, 3, 4}, {5, 6, 7, 8, 9, 10, 11, 12}, {13, 14, 15, 16}]
>>> cm = ClusteringMetrics.from_partitions(ground_truth, predicted)
>>> cm.split_join_distance(normalize=False)
4

For development and testing, this package sets up a Python virtualenv under ./env/ relative to the source tree root.

git clone https://github.com/escherba/clustering-metrics.git
cd clustering-metrics
make test

The above should finish without interruptions and all tests should pass. To generate documentation:

make doc-sources
make doc-html
make doc-publish

This package is under a BSD 3-clause license.