项目代码见 Github：

1.算法介绍

2.代码所用数据

详情参见http://qwone.com/~jason/20Newsgroups/

文件结构

├─doc_classification.py
├─stopwords.txt
├─vocabulary.txt
├─train.data
├─train.label
├─train.map
├─test.data
├─test.label
└─test.map

python代码

需要安装的库:

pandas, liblinearutil

注：Windows平台下 liblinearutil 安装包（32/64）

# doc_classification.py
import pandas as pd
import math
from liblinearutil import *
import time

# 读取数据
def loadOriginData(src='train'):
# train.data
dataSrc = r'%s.data' % src
# train.label
labelSrc = r'%s.label' % src
label = pd.read_table(labelSrc, sep=' ', names=['label'])
# train.map
mapSrc = r'%s.map' % src

# 每个文档拥有的terms
doc2term = {}
# 每个term出现在哪些文档
term2doc = {}
# 每个类别下有哪些docs
cate2docs = {}
# TF值
TF = {}
with open(dataSrc, 'r') as f:
for line in f:
str_docIdx, str_wordIdx, str_cnt = line.split()
docIdx = int(str_docIdx)
wordIdx = int(str_wordIdx)
cnt = int(str_cnt)
# update 数据结构
doc2term.setdefault(docIdx, []).append(wordIdx)
term2doc.setdefault(wordIdx, []).append(docIdx)
TF.setdefault(docIdx, {})[wordIdx] = cnt
# 统计每个类别下有哪些文档
with open(labelSrc, 'r') as f:
for line_index, line in enumerate(f, 1):
labelVal = int(line.strip())
cate2docs.setdefault(labelVal, []).append(line_index)
return TF, doc2term, term2doc, cate2docs, label

# 特征选择
def featureSel(doc2term, term2doc, cate2docs):
# CHI衡量的是特征项ti和类别Cj之间的关联程度, A,B, C, D是四个统计量
CHI_cat2term = {}
# N：total number of documents
N = len(doc2term)
# A + B + C + D = N
# A： term出现在某类别中的文档总数
A = {}
# B: term出现在除某类别外的其他文档数
B = {}
# C: 该类别中不包含term的文档总数
C = {}
# D: 其他类别中不包含term的文档总数
D = {}
DF = {}
# 所有类别
categories = list(cate2docs.keys())
# 停用词词汇表
stopwords = {}
stopwordsSrc = r'stopwords.txt'
with open(stopwordsSrc) as f:
for line in f:
stopwords[line.strip()] = True
# 训练数据数据词汇表
vocSrc = r'vocabulary.txt'
voc = pd.read_table(vocSrc, names=['voc'])
# 保存所有的特征
features = set()
# 计算一个类别标签下各个词的CHI
for category in categories:
# 属于第category类的文档为docs
docs = cate2docs[category]
sumVal = 0
for term in term2doc:
# 如果是停用词, 则将CHI置零
if stopwords.get(voc['voc'][term - 1], False):
CHI_cat2term.setdefault(category, {})[term] = 0
continue
# 属于某类且包含term
AVal = len(set(term2doc[term]).intersection(set(docs)))
# 不属于某类但包含term
BVal = len(term2doc[term]) - AVal
# 属于某类，但不包含term
CVal = len(docs) - AVal
# 不属于某类， 不包含term
DVal = N - AVal - BVal - CVal
CHIVal = N * (AVal * DVal - CVal * BVal)**2 / ((AVal + CVal) * (BVal + DVal) * (AVal + BVal) * (CVal + DVal))
# CHIVal = math.log(AVal * N / ((AVal + CVal) * (AVal + BVal)))
A.setdefault((term, category), AVal)
B.setdefault((term, category), BVal)
C.setdefault((term, category), CVal)
D.setdefault((term, category), DVal)

CHI_cat2term.setdefault(category, {})[term] = CHIVal
DF[term] = AVal + BVal
sumVal += CHIVal
# 选出类别中CHI高于平均值的词
terms = CHI_cat2term[category]
meanVal = sumVal / len(terms)
for term in terms:
if CHI_cat2term[category][term] > meanVal:
features.add(term)
# for feature in features:
# print(voc['voc'][feature])
print('There are %d features in VSM model.\n' % len(features))
return features, DF

def buildSVMData(TF, DF, features, N, label, cate2docs, doc2terms):
isFeatures = dict(zip(features, [True] * len(features)))
categories = list(cate2docs.keys())
# 如果是训练样本， 则计算归一化缩放因子，并返回
# y： label值
y = [0] * N
# x: 稀疏矩阵
x = []
for i in range(N):
x.append({})
for category in categories:
for doc in cate2docs[category]:
# 给y进行标记类别
y[doc - 1] = label.iat[doc - 1, 0]
scale_factor = -100
for term in doc2terms[doc]:
if isFeatures.get(term, False): # 如果term是特征
# TF值
TFVal = TF[doc].get(term, 0)
# TF-IDF值
tf_idf = TFVal * math.log(N / DF[term])
x[doc - 1][term] = tf_idf
# 更新特征最大值
if scale_factor < tf_idf:
scale_factor = tf_idf
alpha = 0
# 按一篇文档中特征词最大的tf-idf, 对该文档中的所有特征词进行归一化
for term in doc2terms[doc]:
if isFeatures.get(term, False): # 如果term是特征
# x[doc - 1][term] = alpha + (1 - alpha) * x[doc - 1][term] / scale_factor
x[doc - 1][term] /= scale_factor
print("Data for SVM has been built.\n")
return x, y

# 计算DF
def getDF(doc2term, term2doc, cate2docs):
DF = {}
for term in term2doc:
DF[term] = len(term2doc[term])
return DF

if __name__ == '__main__':
start = time.time()
# # 主程序
TF, doc2term, term2doc, cate2docs, label = loadOriginData()
# 特征选择
features, DF = featureSel(doc2term, term2doc, cate2docs)
# 读取数据(train.data)
TF, doc2term, term2doc, cate2docs, label = loadOriginData()
# 特征选择
features, DF = featureSel(doc2term, term2doc, cate2docs)
# build SVM model
x, y = buildSVMData(TF, DF, features, len(doc2term), label, cate2docs, doc2term)
# 读取测试数据(test.data)
TF_test, doc2term_test, term2doc_test, cate2docs_test, label_test = loadOriginData(src='test')
DF_test = getDF(doc2term_test, term2doc_test, cate2docs_test)
# TF, DF, features, len(doc2term), label, cate2docs, doc2term, scales)
x_test, y_test = buildSVMData(TF_test, DF_test, features, len(doc2term_test), label_test, cate2docs_test, doc2term_test)

print("处理数据使用了 %s s时间。\n" % (time.time() - start))
# # 调用 liblinear 库进行分类
prob = problem(y, x)
param = parameter('-s 0 -c 4 -B 1')
# 训练
m = train(prob, param)
# 预测test.data
p_label, p_acc, p_vals = predict(y_test, x_test, m, '-b 1')
# 评价
ACC, MSE, SCC = evaluations(y_test, p_label)
print('ACC:\n', ACC)
print('MSE', MSE)
print('SCC', SCC)
# 统计每类中错误率
categoriesErrs = {}
for doc_index, doc_label in enumerate(y_test):
if doc_label != int(p_label[doc_index]):
cateogory = label_test.iat[doc_index, 0]
categoriesErrs.setdefault(cateogory, []).append(doc_index + 1)
# with open('outcome.txt', 'wb') as f:
print("错误分类的样本为：\n")
for categoryErr in categoriesErrs:
numOfErr = len(categoriesErrs[categoryErr])
print('第%d类共 %d样本, 被错分的个数为 %d, 比例为 %f %%.\n' % (categoryErr,len(cate2docs_test[categoryErr]), numOfErr, numOfErr/len(cate2docs_test[categoryErr])))

end = time.time()
print("Total time cost is %s s.\n" % (end - start))