The quality of text-clustering depends mainly on two factors:
Some notion of similarity between the documents you want to cluster. For example, it's easy to distinguish between newsarticles about sports and politics in vector space via tfidf-cosine-distance. It's a lot harder to cluster product-reviews in "good" or "bad" based on this measure.
The clustering method itself. You know how many cluster there'll be? Ok, use kmeans. You don't care about accuracy but want to show a nice tree-structure for navigation of search-results? Use some kind of hierarchical clustering.
There is no text-clustering solution, that would work well under any circumstances. And therefore it's probably not enough to take some clustering software out of the box and throw your data at it.
Having said that, here's some experimental code i used some time ago to play around with text-clustering. The documents are represented as normalized tfidf-vectors and the similarity is measured as cosine distance. The clustering method itself is majorclust.
import sys
from math import log, sqrt
from itertools import combinations
def cosine_distance(a, b):
cos = 0.0
a_tfidf = a["tfidf"]
for token, tfidf in b["tfidf"].iteritems():
if token in a_tfidf:
cos += tfidf * a_tfidf[token]
return cos
def normalize(features):
norm = 1.0 / sqrt(sum(i**2 for i in features.itervalues()))
for k, v in features.iteritems():
features[k] = v * norm
return features
def add_tfidf_to(documents):
tokens = {}
for id, doc in enumerate(documents):
tf = {}
doc["tfidf"] = {}
doc_tokens = doc.get("tokens", [])
for token in doc_tokens:
tf[token] = tf.get(token, 0) + 1
num_tokens = len(doc_tokens)
if num_tokens > 0:
for token, freq in tf.iteritems():
tokens.setdefault(token, []).append((id, float(freq) / num_tokens))
doc_count = float(len(documents))
for token, docs in tokens.iteritems():
idf = log(doc_count / len(docs))
for id, tf in docs:
tfidf = tf * idf
if tfidf > 0:
documents[id]["tfidf"][token] = tfidf
for doc in documents:
doc["tfidf"] = normalize(doc["tfidf"])
def choose_cluster(node, cluster_lookup, edges):
new = cluster_lookup[node]
if node in edges:
seen, num_seen = {}, {}
for target, weight in edges.get(node, []):
seen[cluster_lookup[target]] = seen.get(
cluster_lookup[target], 0.0) + weight
for k, v in seen.iteritems():
num_seen.setdefault(v, []).append(k)
new = num_seen[max(num_seen)][0]
return new
def majorclust(graph):
cluster_lookup = dict((node, i) for i, node in enumerate(graph.nodes))
count = 0
movements = set()
finished = False
while not finished:
finished = True
for node in graph.nodes:
new = choose_cluster(node, cluster_lookup, graph.edges)
move = (node, cluster_lookup[node], new)
if new != cluster_lookup[node] and move not in movements:
movements.add(move)
cluster_lookup[node] = new
finished = False
clusters = {}
for k, v in cluster_lookup.iteritems():
clusters.setdefault(v, []).append(k)
return clusters.values()
def get_distance_graph(documents):
class Graph(object):
def __init__(self):
self.edges = {}
def add_edge(self, n1, n2, w):
self.edges.setdefault(n1, []).append((n2, w))
self.edges.setdefault(n2, []).append((n1, w))
graph = Graph()
doc_ids = range(len(documents))
graph.nodes = set(doc_ids)
for a, b in combinations(doc_ids, 2):
graph.add_edge(a, b, cosine_distance(documents[a], documents[b]))
return graph
def get_documents():
texts = [
"foo blub baz",
"foo bar baz",
"asdf bsdf csdf",
"foo bab blub",
"csdf hddf kjtz",
"123 456 890",
"321 890 456 foo",
"123 890 uiop",
]
return [{"text": text, "tokens": text.split()}
for i, text in enumerate(texts)]
def main(args):
documents = get_documents()
add_tfidf_to(documents)
dist_graph = get_distance_graph(documents)
for cluster in majorclust(dist_graph):
print "========="
for doc_id in cluster:
print documents[doc_id]["text"]
if __name__ == '__main__':
main(sys.argv)
For real applications, you would use a decent tokenizer, use integers instead of token-strings and don't calc a O(n^2) distance-matrix...
