Agglomerative clustering using Scikit-Learn (with a custom distance metric)

hclustering.py

"""
Hierarchial Clustering.
The goal of gist is to show to use scikit-learn to perform agglomerative clustering when:
1. There is a need for a custom distance metric (like levenshtein distance)
2. Use the distance in sklearn's API.

Adapted from: sklearn's FAQ.
http://scikit-learn.org/stable/faq.html
"""

# Using editdistance package to calculate the levenhstein distance between strings.
import editdistance as edist

import numpy as np
from sklearn.cluster import AgglomerativeClustering
from sklearn.metrics.pairwise import pairwise_distances

# Simple data (sample genome string).
data = ["ACCTCCTAGAAG", "ACCTACTAGAAGTT", "GAATATTAGGCCGA"]

# Custom distance metric and use editdistance.
def lev_metric(x, y):
    return int(edist.eval(data[int(x[0])], data[int(y[0])]))

# Reshape the data.
X = np.arange(len(data)).reshape(-1, 1)
print(X.shape)

# Calculate pairwise distances with the new metric.
m = pairwise_distances(X, X, metric=lev_metric)
print(m)

# Perform agglomerative clustering.
# The affinity is precomputed (since the distance are precalculated).
# Use an 'average' linkage. Use any other apart from  'ward'.
agg = AgglomerativeClustering(n_clusters=2, affinity='precomputed',
                              linkage='average')

# Use the distance matrix directly.
u = agg.fit_predict(m)
print(u)