Collection of NLP clustering algorithms

Currently implemented are:

* KMeans with cosine distances
* Chinese Whispers

Installation

Its not (yet) uploaded to PyPI, so installation has to be done from source:

git clone https://github.com/tttthomasssss/clustering.pip
cd clustering/
pip install -e .

Cosine-Means a.k.a. Spherical KMeans or just KMeans with cosine distance

Because sklearns KMeans implementation doesn't allow different distance metrics.

Kind of an intermediary between standard KMeans and SpectralClustering with affinity='cosine' (SpectralClustering first performs a low-dimensional embedding of the affinity matrix and clusters it, whereas CosineMeans works directly on the data matrix).

Usage:

from cosine_means import CosineMeans
from sklearn.datasets import make_classification
X, _ = make_classification(n_samples=100, n_features=2, n_informative=1, n_redundant=0, n_clusters_per_class=1)

cmeans = CosineMeans(n_clusters=2)
cmeans.fit(X)

print(cmeans.cluster_centers_)

Chinese Whispers

Note: requires networkX

see Biemann (2006) - Chinese Whispers - an Efficient Graph Clustering Algorithm and its Application to Natural Language Processing Problems

Usage:

import networkx as nx

# Create random adjacency matrix
rnd = np.random.RandomState(seed=1105)
X = rnd.binomial(n=1, p=0.2, size=(10, 10))

# Create graph from adjacency matrix
G = nx.from_numpy_matrix(np.asmatrix(X))

# Cluster graph
cw = ChineseWhispers(max_iter=5)
y = cw.fit_predict(G)

print(y)