mirror of
https://gitee.com/TheAlgorithms/Statistical-Learning-Method_Code.git
synced 2024-12-22 12:48:55 +08:00
Add LDA.py
This commit is contained in:
parent
20cf886342
commit
1e1898b5fa
187
LDA/LDA.py
Normal file
187
LDA/LDA.py
Normal file
@ -0,0 +1,187 @@
|
||||
#coding=utf-8
|
||||
#Author:Harold
|
||||
#Date:2021-1-27
|
||||
#Email:zenghr_zero@163.com
|
||||
|
||||
'''
|
||||
数据集:bbc_text
|
||||
数据集数量:2225
|
||||
-----------------------------
|
||||
运行结果:
|
||||
话题数:5
|
||||
原始话题:'tech', 'business', 'sport', 'entertainment', 'politics'
|
||||
生成话题:
|
||||
1:'said game england would time first back play last good'
|
||||
2:'said year would economy growth also economic bank government could'
|
||||
3:'said year games sales company also market last firm 2004'
|
||||
4:'film said music best also people year show number digital'
|
||||
5:'said would people government labour election party blair could also'
|
||||
运行时长:7620.51s
|
||||
'''
|
||||
|
||||
import numpy as np
|
||||
import pandas as pd
|
||||
import string
|
||||
from nltk.corpus import stopwords
|
||||
import time
|
||||
|
||||
|
||||
#定义加载数据的函数
|
||||
def load_data(file, K):
|
||||
'''
|
||||
INPUT:
|
||||
file - (str) 数据文件的路径
|
||||
K - (int) 设定的话题数
|
||||
|
||||
OUTPUT:
|
||||
org_topics - (list) 原始话题标签列表
|
||||
text - (list) 文本列表
|
||||
words - (list) 单词列表
|
||||
alpha - (list) 话题概率分布,模型超参数
|
||||
beta - (list) 单词概率分布,模型超参数
|
||||
|
||||
'''
|
||||
df = pd.read_csv(file) #读取文件
|
||||
org_topics = df['category'].unique().tolist() #保存文本原始的话题标签
|
||||
M = df.shape[0] #文本数
|
||||
alpha = np.zeros(K) #alpha是LDA模型的一个超参数,是对话题概率的预估计,这里取文本数据中各话题的比例作为alpha值,实际可以通过模型训练得到
|
||||
beta = np.zeros(1000) #beta是LDA模型的另一个超参数,是词汇表中单词的概率分布,这里取各单词在所有文本中的比例作为beta值,实际也可以通过模型训练得到
|
||||
#计算各话题的比例作为alpha值
|
||||
for k, topic in enumerate(org_topics):
|
||||
alpha[k] = df[df['category'] == topic].shape[0] / M
|
||||
df.drop('category', axis=1, inplace=True)
|
||||
n = df.shape[0] #n为文本数量
|
||||
text = []
|
||||
words = []
|
||||
for i in df['text'].values:
|
||||
t = i.translate(str.maketrans('', '', string.punctuation)) #去除文本中的标点符号
|
||||
t = [j for j in t.split() if j not in stopwords.words('english')] #去除文本中的停止词
|
||||
t = [j for j in t if len(j) > 3] #长度小于等于3的单词大多是无意义的,直接去除
|
||||
text.append(t) #将处理后的文本保存到文本列表中
|
||||
words.extend(set(t)) #将文本中所包含的单词保存到单词列表中
|
||||
words = list(set(words)) #去除单词列表中的重复单词
|
||||
words_cnt = np.zeros(len(words)) #用来保存单词的出现频次
|
||||
#循环计算words列表中各单词出现的词频
|
||||
for i in range(len(text)):
|
||||
t = text[i] #取出第i条文本
|
||||
for w in t:
|
||||
ind = words.index(w) #取出第i条文本中的第t个单词在单词列表中的索引
|
||||
words_cnt[ind] += 1 #对应位置的单词出现频次加一
|
||||
sort_inds = np.argsort(words_cnt)[::-1] #对单词出现频次降序排列后取出其索引值
|
||||
words = [words[ind] for ind in sort_inds[:1000]] #将出现频次前1000的单词保存到words列表
|
||||
#去除文本text中不在词汇表words中的单词
|
||||
for i in range(len(text)):
|
||||
t = []
|
||||
for w in text[i]:
|
||||
if w in words:
|
||||
ind = words.index(w)
|
||||
t.append(w)
|
||||
beta[ind] += 1 #统计各单词在文本中的出现频次
|
||||
text[i] = t
|
||||
beta /= np.sum(beta) #除以文本的总单词数得到各单词所占比例,作为beta值
|
||||
return org_topics, text, words, alpha, beta
|
||||
|
||||
|
||||
#定义潜在狄利克雷分配函数,采用收缩的吉布斯抽样算法估计模型的参数theta和phi
|
||||
def do_lda(text, words, alpha, beta, K, iters):
|
||||
'''
|
||||
INPUT:
|
||||
text - (list) 文本列表
|
||||
words - (list) 单词列表
|
||||
alpha - (list) 话题概率分布,模型超参数
|
||||
beta - (list) 单词概率分布,模型超参数
|
||||
K - (int) 设定的话题数
|
||||
iters - (int) 设定的迭代次数
|
||||
|
||||
OUTPUT:
|
||||
theta - (array) 话题的条件概率分布p(zk|dj),这里写成p(zk|dj)是为了和PLSA模型那一章的符号统一一下,方便对照着看
|
||||
phi - (array) 单词的条件概率分布p(wi|zk)
|
||||
|
||||
'''
|
||||
M = len(text) #文本数
|
||||
V = len(words) #单词数
|
||||
N_MK = np.zeros((M, K)) #文本-话题计数矩阵
|
||||
N_KV = np.zeros((K, V)) #话题-单词计数矩阵
|
||||
N_M = np.zeros(M) #文本计数向量
|
||||
N_K = np.zeros(K) #话题计数向量
|
||||
Z_MN = [] #用来保存每条文本的每个单词所在位置处抽样得到的话题
|
||||
#算法20.2的步骤(2),对每个文本的所有单词抽样产生话题,并进行计数
|
||||
for m in range(M):
|
||||
zm = []
|
||||
t = text[m]
|
||||
for n, w in enumerate(t):
|
||||
v = words.index(w)
|
||||
z = np.random.randint(K)
|
||||
zm.append(z)
|
||||
N_MK[m, z] += 1
|
||||
N_M[m] += 1
|
||||
N_KV[z, v] += 1
|
||||
N_K[z] += 1
|
||||
Z_MN.append(zm)
|
||||
#算法20.2的步骤(3),多次迭代进行吉布斯抽样
|
||||
for i in range(iters):
|
||||
print('{}/{}'.format(i+1, iters))
|
||||
for m in range(M):
|
||||
t = text[m]
|
||||
for n, w in enumerate(t):
|
||||
v = words.index(w)
|
||||
z = Z_MN[m][n]
|
||||
N_MK[m, z] -= 1
|
||||
N_M[m] -= 1
|
||||
N_KV[z][v] -= 1
|
||||
N_K[z] -= 1
|
||||
p = [] #用来保存对K个话题的条件分布p(zi|z_i,w,alpha,beta)的计算结果
|
||||
sums_k = 0
|
||||
for k in range(K):
|
||||
p_zk = (N_KV[k][v] + beta[v]) * (N_MK[m][k] + alpha[k]) #话题zi=k的条件分布p(zi|z_i,w,alpha,beta)的分子部分
|
||||
sums_v = 0
|
||||
sums_k += N_MK[m][k] + alpha[k] #累计(nmk + alpha_k)在K个话题上的和
|
||||
for t in range(V):
|
||||
sums_v += N_KV[k][t] + beta[t] #累计(nkv + beta_v)在V个单词上的和
|
||||
p_zk /= sums_v
|
||||
p.append(p_zk)
|
||||
p = p / sums_k
|
||||
p = p / np.sum(p) #对条件分布p(zi|z_i,w,alpha,beta)进行归一化,保证概率的总和为1
|
||||
new_z = np.random.choice(a=K, p=p) #根据以上计算得到的概率进行抽样,得到新的话题
|
||||
Z_MN[m][n] = new_z #更新当前位置处的话题为上面抽样得到的新话题
|
||||
#更新计数
|
||||
N_MK[m, new_z] += 1
|
||||
N_M[m] += 1
|
||||
N_KV[new_z, v] += 1
|
||||
N_K[new_z] += 1
|
||||
#算法20.2的步骤(4),利用得到的样本计数,估计模型的参数theta和phi
|
||||
theta = np.zeros((M, K))
|
||||
phi = np.zeros((K, V))
|
||||
for m in range(M):
|
||||
sums_k = 0
|
||||
for k in range(K):
|
||||
theta[m, k] = N_MK[m][k] + alpha[k] #参数theta的分子部分
|
||||
sums_k += theta[m, k] #累计(nmk + alpha_k)在K个话题上的和,参数theta的分母部分
|
||||
theta[m] /= sums_k #计算参数theta
|
||||
for k in range(K):
|
||||
sums_v = 0
|
||||
for v in range(V):
|
||||
phi[k, v] = N_KV[k][v] + beta[v] #参数phi的分子部分
|
||||
sums_v += phi[k][v] #累计(nkv + beta_v)在V个单词上的和,参数phi的分母部分
|
||||
phi[k] /= sums_v #计算参数phi
|
||||
return theta, phi
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
K = 5 #设定话题数为5
|
||||
org_topics, text, words, alpha, beta = load_data('bbc_text.csv', K) #加载数据
|
||||
print('Original Topics:')
|
||||
print(org_topics) #打印原始的话题标签列表
|
||||
start = time.time() #保存开始时间
|
||||
iters = 10 #为了避免运行时间过长,这里只迭代10次,实际上10次是不够的,要迭代足够的次数保证吉布斯抽样进入燃烧期,这样得到的参数才能尽可能接近样本的实际概率分布
|
||||
theta, phi = do_lda(text, words, alpha, beta, K, iters) #LDA的吉布斯抽样
|
||||
#打印出每个话题zk条件下出现概率最大的前10个单词,即P(wi|zk)在话题zk中最大的10个值对应的单词,作为对话题zk的文本描述
|
||||
for k in range(K):
|
||||
sort_inds = np.argsort(phi[k])[::-1] #对话题zk条件下的P(wi|zk)的值进行降序排列后取出对应的索引值
|
||||
topic = [] #定义一个空列表用于保存话题zk概率最大的前10个单词
|
||||
for i in range(10):
|
||||
topic.append(words[sort_inds[i]])
|
||||
topic = ' '.join(topic) #将10个单词以空格分隔,构成对话题zk的文本表述
|
||||
print('Topic {}: {}'.format(k+1, topic)) #打印话题zk
|
||||
end = time.time()
|
||||
print('Time:', end-start)
|
Loading…
Reference in New Issue
Block a user