scikit-learn kmeans实现文本聚类

kmeans 无监督的学习方法。需要根据实际业务需要确定K值。

• 加载数据集

%matplotlib inline
import matplotlib.pyplot as plt
import numpy as np

from time import time
from sklearn.datasets import load_files

print("loading documents ...")
t = time()
docs = load_files('datasets/clustering/data')
print("summary: {0} documents in {1} categories.".format(
    len(docs.data), len(docs.target_names)))
print("done in {0} seconds".format(time() - t))

• 文档向量化

from sklearn.feature_extraction.text import TfidfVectorizer

max_features = 20000
print("vectorizing documents ...")
t = time()
vectorizer = TfidfVectorizer(max_df=0.4, 
                             min_df=2, 
                             max_features=max_features, 
                             encoding='latin-1')
X = vectorizer.fit_transform((d for d in docs.data))
print("n_samples: %d, n_features: %d" % X.shape)
print("number of non-zero features in sample [{0}]: {1}".format(
    docs.filenames[0], X[0].getnnz()))
print("done in {0} seconds".format(time() - t))

• 聚类

from sklearn.cluster import KMeans

print("clustering documents ...")
t = time()
n_clusters = 4
kmean = KMeans(n_clusters=n_clusters, 
               max_iter=100,
               tol=0.01,
               verbose=1,
               n_init=3)
kmean.fit(X);
print("kmean: k={}, cost={}".format(n_clusters, int(kmean.inertia_)))
print("done in {0} seconds".format(time() - t))

• 分类过程中权重高的10个词

from __future__ import print_function

print("Top terms per cluster:")

order_centroids = kmean.cluster_centers_.argsort()[:, ::-1]

terms = vectorizer.get_feature_names()
for i in range(n_clusters):
    print("Cluster %d:" % i, end='')
    for ind in order_centroids[i, :10]:
        print(' %s' % terms[ind], end='')
    print()

如何确定聚类结果的好坏呢？

主要有一下几个属性：

Adjust Rand Index:衡量两个序列相似性的算法，优点是针对两个随机序列，值是负数或者接近于0，如果是两个结构相同的序列，值接近于1，对类别标签不敏感。

from sklearn import metrics

label_true = np.random.randint(1, 4, 6)
label_pred = np.random.randint(1, 4, 6)
print("Adjusted Rand-Index for random sample: %.3f"
      % metrics.adjusted_rand_score(label_true, label_pred))
label_true = [1, 1, 3, 3, 2, 2]
label_pred = [3, 3, 2, 2, 1, 1]
print("Adjusted Rand-Index for same structure sample: %.3f"
      % metrics.adjusted_rand_score(label_true, label_pred))

齐次性homogeneity和完整性completeness

齐次性表示一个聚类元素只由一种类别的元素组成。

完整性表示给定已经标记的类别，全部分配到一个聚类里。

齐次性和完整性是一个互补的关系，两个指标综合起来称为V-measure分数。

from sklearn import metrics

label_true = [1, 1, 2, 2]
label_pred = [2, 2, 1, 1]
print("Homogeneity score for same structure sample: %.3f"
      % metrics.homogeneity_score(label_true, label_pred))
label_true = [1, 1, 2, 2]
label_pred = [0, 1, 2, 3]
print("Homogeneity score for each cluster come from only one class: %.3f"
      % metrics.homogeneity_score(label_true, label_pred))
label_true = [1, 1, 2, 2]
label_pred = [1, 2, 1, 2]
print("Homogeneity score for each cluster come from two class: %.3f"
      % metrics.homogeneity_score(label_true, label_pred))
label_true = np.random.randint(1, 4, 6)
label_pred = np.random.randint(1, 4, 6)
print("Homogeneity score for random sample: %.3f"
      % metrics.homogeneity_score(label_true, label_pred))


from sklearn import metrics

label_true = [1, 1, 2, 2]
label_pred = [2, 2, 1, 1]
print("Completeness score for same structure sample: %.3f"
      % metrics.completeness_score(label_true, label_pred))
label_true = [0, 1, 2, 3]
label_pred = [1, 1, 2, 2]
print("Completeness score for each class assign to only one cluster: %.3f"
      % metrics.completeness_score(label_true, label_pred))
label_true = [1, 1, 2, 2]
label_pred = [1, 2, 1, 2]
print("Completeness score for each class assign to two class: %.3f"
      % metrics.completeness_score(label_true, label_pred))
label_true = np.random.randint(1, 4, 6)
label_pred = np.random.randint(1, 4, 6)
print("Completeness score for random sample: %.3f"
      % metrics.completeness_score(label_true, label_pred))

from sklearn import metrics

label_true = [1, 1, 2, 2]
label_pred = [2, 2, 1, 1]
print("V-measure score for same structure sample: %.3f"
      % metrics.v_measure_score(label_true, label_pred))
label_true = [0, 1, 2, 3]
label_pred = [1, 1, 2, 2]
print("V-measure score for each class assign to only one cluster: %.3f"
      % metrics.v_measure_score(label_true, label_pred))
print("V-measure score for each class assign to only one cluster: %.3f"
      % metrics.v_measure_score(label_pred, label_true))
label_true = [1, 1, 2, 2]
label_pred = [1, 2, 1, 2]
print("V-measure score for each class assign to two class: %.3f"
      % metrics.v_measure_score(label_true, label_pred))

轮廓系数

from sklearn import metrics

labels = docs.target
print("Homogeneity: %0.3f" % metrics.homogeneity_score(labels, kmean.labels_))
print("Completeness: %0.3f" % metrics.completeness_score(labels, kmean.labels_))
print("V-measure: %0.3f" % metrics.v_measure_score(labels, kmean.labels_))
print("Adjusted Rand-Index: %.3f"
      % metrics.adjusted_rand_score(labels, kmean.labels_))
print("Silhouette Coefficient: %0.3f"
      % metrics.silhouette_score(X, kmean.labels_, sample_size=1000))

轮廓系数可以在不需要已标记的数据集的前提下，对聚类算法的性能进行评估。

a:一个样本与其所在相同聚类的平均距离

b:一个样本预期距离最近的下一个聚类里的点的平均距离。

轮廓系数s = (b-a)/max(a,b) 值介于[-1, 1]之间，-1表示完全错误的聚类，1表示完美的聚类，0表示聚类重叠