基于KNN的文本分类实战

本文讲述如何使用scikit-learn的KNN工具对文本进行分类。

关于KNN

---

K-近邻算法，简称KNN（k-Nearest Neighbor），是一个相当简单的分类/预测算法。其主要思想就是，选取与待分类/预测数据的最相似的K个训练数据，通过对这K个数据的结果或者分类标号取平均、取众数等方法得到待分类/预测数据的结果或者分类标号。

关于KNN，笔者在浅入浅出：K近邻算法有较为详细的介绍。

数据集介绍

---

数据集是有8个分类的文本数据集，使用了结巴分词对每个文本分词，每个单词当作特征，再利用二元词串构造更多特征，然后去掉停用词，去掉出现次数太多和太少的特征，得到了19630个特征。取1998个样本用于训练，509个用于测试。基于词袋模型的思路将每个文本转换为向量，训练集和测试集分别转换为矩阵，并用python numpy模块将其保存为npy格式。

在 https://github.com/someus/dataset 下载text-classification.7z，解压后导入数据：

$ ls
test_data.npy  test_labels.npy  training_data.npy  training_labels.npy  
$ ipython
>>> import numpy as np
>>> training_data = np.load("training_data.npy")
>>> training_data.shape
(1998, 19630)
>>> training_labels = np.load("training_labels.npy")
>>> training_labels
array([6, 6, 6, ..., 2, 2, 2])  
>>> training_labels.shape
(1998,)
>>> test_data = np.load("test_data.npy")
>>> test_data.shape
(509, 19630)
>>> test_labels = np.load("test_labels.npy")
>>> test_labels.shape
(509,)

如何找一样本的最近k个邻居

---

方法1：

>>> from sklearn.neighbors import NearestNeighbors
>>> nbrs = NearestNeighbors(n_neighbors=6, algorithm='ball_tree')
>>> nbrs.fit(training_data)  # 构造BallTree，可以快速找出6个最近邻居，原理待学习
NearestNeighbors(algorithm='ball_tree', leaf_size=30, metric='minkowski',
         metric_params=None, n_neighbors=6, p=2, radius=1.0)
>>> distances, indices = nbrs.kneighbors(test_data[0])  # 找training_data中离样本test_data[0]的最近的6个样本
>>> indices  # 6个最近样本，每个值是指在training_data中的第几个样本
array([[500, 294,  62, 802, 732, 703]])
>>> distances  # 对应的距离
array([[ 13.37908816,  13.60147051,  13.60147051,  13.60147051,
         13.60147051,  13.6381817 ]])

也可以依次找出多个测试样本的最近的6个训练样本：

>>> distances, indices = nbrs.kneighbors(test_data[0:2])
>>> indices
array([[ 500,  294,   62,  802,  732,  703],
       [  62,  294,  636, 1945,  802, 1091]])
>>> distances
array([[ 13.37908816,  13.60147051,  13.60147051,  13.60147051,
         13.60147051,  13.6381817 ],
       [  7.93725393,   7.93725393,   8.1240384 ,   8.36660027,
          8.54400375,   8.54400375]])

方法2：

>>> from sklearn.neighbors import BallTree
>>> bt = BallTree(training_data, metric='euclidean')
>>> distances, indices = bt.query(test_data[0], k=6)                  
>>> indices
array([[500,  62, 802, 294, 732, 703]])
>>> distances
array([[ 13.37908816,  13.60147051,  13.60147051,  13.60147051,
         13.60147051,  13.6381817 ]])

基于KNN的文本分类

---

令k=6：

>>> from sklearn.neighbors import KNeighborsClassifier
>>> knn = KNeighborsClassifier(n_neighbors=6, metric='euclidean')
>>> knn.fit(training_data, training_labels) # 训练
KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='euclidean',
           metric_params=None, n_neighbors=6, p=2, weights='uniform')
>>> predict_labels = knn.predict(test_data) # 预测
>>> sum(predict_labels == test_labels)
230
>>> 230./509  # 正确率
0.4518664047151277

令k=20：

>>> from sklearn.neighbors import KNeighborsClassifier
>>> knn = KNeighborsClassifier(n_neighbors=20, metric='euclidean')
>>> knn.fit(training_data, training_labels) # 训练
KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='euclidean',
           metric_params=None, n_neighbors=20, p=2, weights='uniform')
>>> predict_labels = knn.predict(test_data) # 预测
>>> sum(predict_labels == test_labels)
276  # 效果比k=6时提升了一些
>>> 276./509   # 正确率
0.5422396856581533

这个正确率并不高。在基于贝叶斯的文本分类实战中笔者使用了多项式贝叶斯对同样的数据集进行分类，正确率达到近90%。

做个优化

---

我们将每个样本归一化，看看效果。

先写一个归一化工具（mytools.py）：

# !/usr/bin/env python
# -*- encoding:utf-8 -*-

import numpy as np

def uniformization(X):
    if X.ndim != 2:
        return None
    X2 = X.copy()
    X2 = X2.astype(float)
    rows = X2.shape[0]
    for i in xrange(0, rows):
        sum_of_squares = sum(X2[i, :]**2)
        if sum_of_squares == 0: continue
        sqrt_sum_of_squares = sum_of_squares**0.5
        X2[i, :] = X2[i, :] / sqrt_sum_of_squares
    return X2 

if __name__ == '__main__':
    arr = np.array([[1,2,3],[4,5,6],[0,0,0]])
    print uniformization(arr)

运行结果如下：

[[ 0.26726124  0.53452248  0.80178373]
 [ 0.45584231  0.56980288  0.68376346]
 [ 0.          0.          0.

处理原始数据集，生成新的数据：

>>> from mytools import uniformization
>>> new_training_data = uniformization(training_data)
>>> new_test_data = uniformization(test_data)

令k=6：

>>> knn = KNeighborsClassifier(n_neighbors=6, metric='euclidean')
>>> knn.fit(new_training_data, training_labels) # 使用新数据训练
KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='euclidean',
           metric_params=None, n_neighbors=6, p=2, weights='uniform')
>>> predict_labels = knn.predict(new_test_data) # 预测
>>> sum(predict_labels == test_labels)
294  # 由230提升到294
>>> 294./509  # 正确率有提升
0.5776031434184676

令k=20：

>>> from sklearn.neighbors import KNeighborsClassifier
>>> knn = KNeighborsClassifier(n_neighbors=20, metric='euclidean')
>>> knn.fit(new_training_data, training_labels)  # 使用新数据训练
KNeighborsClassifier(algorithm='auto', leaf_size=30, metric='euclidean',
           metric_params=None, n_neighbors=20, p=2, weights='uniform')
>>> predict_labels = knn.predict(new_test_data)  # 预测
>>> sum(predict_labels == test_labels)
314  # 由276提升到314
>>> 314./509  # 正确率有提升
0.6168958742632613

可以看到，归一化后，预测分类的正确率提升很多。

参考

---

1.6. Nearest Neighbors
sklearn.neighbors.KNeighborsClassifier