flexible-clustering-tree

---

What’s this?

In the context of clustering task, flexible-clustering-tree provides you easy and controllable clustering framework.

|image0|

Background

Let’s suppose, you have huge data. You’d like to observe data as easy as possible.

Hierarchical clustering is a way to see clustering tree. However, hierarchical clustering tends to fall into local optimization.

So, you need other clustering method. But at the same time, you wanna observe your data with tree structure style.

here, flexible-clustering-tree could give you simple way from data into tree viewer(d3 based)

You could set any kinds of clustering algorithm such as Kmeans, DBSCAN, Spectral-Clustering.

Multi feature and Multi clustering

During making a tree, you might want use various kind of clustering algorithm. For example, you use Kmeans for the 1st later of a tree, and DBSCAN for the 2nd layer of a tree.

And you might also use various kind of feature type depending on a layer of a tree. For example, in the context of document clustering, “title” of news for the 1st layer, and “whole text” for the 2nd layer.

The example below, this is a clustering tree of 20-news data set.

• 1st layer(red highlight) is done with HDBSCAN clustering, and feature is dense vector of Subject text, which is converted by word2vec model.
• 2nd layer(blue highlight) is done with Kmeans, and feature is sparse vector of whole text(BOW).

You could design your clustering tree as you want!

|image1|

Both are possible flexible-clustering-tree!

Contribution

• Easy interface(scikit-learn way) from data(feature matrix) into a tree viewer
• Possible to make various clustering algorithms ensemble
• Possible to set various feature types

How to use?

.. code:: python

from flexible_clustering_tree import FeatureMatrixObject, MultiFeatureMatrixObject from flexible_clustering_tree import ClusteringOperator, MultiClusteringOperator from flexible_clustering_tree import FlexibleClustering import numpy import codecs

set feature matrix

an input of 1st layer is list of string

input_string = ['d'] * 10 + ['e'] * 10 + ['c'] * 10 + ['a'] * 10 + ['b'] * 10 + ['f'] * 50 f_obj_1st = FeatureMatrixObject(0, feature_strings=input_string)

an input of 2nd layer is the dense matrix (100, 300)

f_obj_2nd = FeatureMatrixObject(1, matrix_object=numpy.random.rand(100, 300))

an input of 3rd layer is the dense matrix (100, 50)

f_obj_3rd = FeatureMatrixObject(2, matrix_object=numpy.random.rand(100, 50)) dict_index2label = {i: "label-{}".format(i) for i in range(0, 100)} multi_feature_matrix = MultiFeatureMatrixObject( [f_obj_1st, f_obj_2nd, f_obj_3rd], dict_index2label=dict_index2label )

set clustering operation

from sklearn.cluster.k_means_ import KMeans from hdbscan.hdbscan_ import HDBSCAN from flexible_clustering_tree import StringAggregation c_operation_1st = ClusteringOperator(0, -1, StringAggregation()) c_operation_2nd = ClusteringOperator(1, 10, KMeans(10)) c_operation_3rd = ClusteringOperator(2, -1, HDBSCAN()) multi_clustering = MultiClusteringOperator([c_operation_1st, c_operation_2nd])

run flexible clustering

clustering_runner = FlexibleClustering(max_depth=5) index2cluster_no = clustering_runner.fit_transform(multi_feature_matrix, multi_clustering) html = clustering_runner.clustering_tree.to_html()

output to html

with codecs.open("out.html", "w", "utf-8") as f: f.write(html)

you're also able to generate tables via Pandas.

import pandas table_objects = clustering_runner.clustering_tree.to_objects() print(pandas.DataFrame(table_objects['cluster_information'])) print(pandas.DataFrame(table_objects['leaf_information']))

The output of pandas table is below.

The relation-table of clusters is in cluster_information.

::

   cluster_id  parent_id  depth_level  clustering_method

0 0 -1 1 StringAggregation 1 1 -1 1 StringAggregation 2 2 -1 1 StringAggregation 3 3 -1 1 StringAggregation 4 4 -1 1 StringAggregation 5 5 -1 1 StringAggregation 6 6 5 2 KMeans 7 7 5 2 KMeans 8 8 5 2 KMeans 9 9 5 2 KMeans 10 10 5 2 KMeans 11 11 5 2 KMeans 12 12 5 2 KMeans 13 13 5 2 KMeans 14 14 5 2 KMeans 15 15 5 2 KMeans

The relation-table of leaf nodes is in leaf_information.

::

   leaf_id  cluster_id     label  args

0 0 0 label-0 None 1 1 0 label-1 None 2 2 0 label-2 None 3 3 0 label-3 None 4 4 0 label-4 None .. ... ... ... ... 95 95 14 label-95 None 96 96 8 label-96 None 97 97 13 label-97 None 98 98 10 label-98 None 99 99 12 label-99 None [100 rows x 4 columns]

You could see examples at /examples.

setup

.. code:: bash

pip install flexible_clustering_tree

or close this repository

.. code:: bash

python setup.py install

For Developers

Environment

• Python >= 3.x

Dev/Test environment by Docker

You build dev/test environment with Docker container. Here is simple procedure,

1. build docker image
2. start docker container
3. run test in the container

.. code:: bash

$ cd tests $ docker-compose build $ docker-compose up $ docker run --name test-container -v pwd:/codes/flexible-clustering-tree/ -dt tests_dev_env $ docker exec -it test-container python /codes/flexible-clustering-tree/setup.py test

If you’re using pycharm professional edition, you could call a docker-compose file as Python interpreter.

.. |image0| image:: https://user-images.githubusercontent.com/1772712/47308081-9980cd00-d66b-11e8-98c0-a275db021cd7.gif .. |image1| image:: https://user-images.githubusercontent.com/1772712/47308468-abaf3b00-d66c-11e8-9a08-26facc39e80e.png