Sklearn AffinityPropagation 内存错误

Question

我想我已经知道我的答案了，但是有很多比我更聪明、更有经验的人所以我想问一下。

当我试图让我的 hash_matrix (<class 'scipy.sparse.csr.csr_matrix'>) 适合 AffinityPropagation 时，我 运行 变成了 MemoryError。它仅在 10,000 个样本上失败，这在我的实际数据集范围内相对较小。

我的问题：我喜欢 AffinityPropagation 在较小的数据集上看到的结果，但除非我能够将其应用到较大的数据集，否则它毫无用处。

我的问题：是否正在尝试将 AffinityPropagation 应用于成百上千的项目，这在标准笔记本电脑上不太可能发生？

我学到了什么：

1. AffinityPropagation does not support partial_fit 和增量学习。
2. time complexity 是 AffinityPropagation
的主要缺点
3. Affinity Propagation [is] most appropriate for small to medium sized datasets.

抛出的错误：

Traceback (most recent call last):
  File "C:/Users/my.name/Documents/my files/Programs/clustering_test/SOexample.py", line 68, in <module>
    aff.fit(hash_matrix)
  File "C:\Python34\lib\site-packages\sklearn\cluster\affinity_propagation_.py", line 301, in fit
    copy=self.copy, verbose=self.verbose, return_n_iter=True)
  File "C:\Python34\lib\site-packages\sklearn\cluster\affinity_propagation_.py", line 105, in affinity_propagation
    S += ((np.finfo(np.double).eps * S + np.finfo(np.double).tiny * 100) *
MemoryError

完整的工作代码示例：

import pandas as pd
import numpy as np

from nltk.stem import PorterStemmer

from sklearn.feature_extraction.text import HashingVectorizer
from sklearn import cluster

data = ['10 news headlines', '3 current events in the news today',
 '5 day break in new york', '7 breaking news', '7 news breaking news',
 '7 news headlines', '7 news online', '7 news today', 'america current news',
 'america new york time', 'america news latest', 'america news online',
 'america news paper', 'america news today', 'america recent news',
 'american news channel', 'american news channels', 'any news today',
 'article about new york', 'article about new york city',
 'article in newspaper today', 'article news today', 'article today news',
 'articles about new york', 'articles on new york', 'articles usa',
 'best news channel', 'best news homepage', 'best newspaper websites',
 'big news stories', 'big news stories of 2013', 'break in new york',
 'break news today', 'break to new york', 'breaking cnn news',
 'breaking entertainment news', 'breaking global news', 'breaking headlines',
 'breaking international news', 'breaking international news today',
 'breaking latest news', 'breaking nation news', 'breaking new cnn',
 'breaking new for today', 'breaking new of today', 'breaking new today',
 'breaking news', 'breaking news america today',
 'breaking news and top stories', 'breaking news around the world',
 'breaking news around the world today', 'breaking news brooklyn',
 'breaking news cnn', 'breaking news cnn alerts', 'breaking news cnn live',
 'world important news today', 'world latest breaking news',
 'world latest news', 'world latest news headlines', 'world latest news today',
 'world latest news update', 'world latest news updates', 'world new headlines',
 'world new now', 'world new today', 'world news', 'world news articles',
 'world news articles today', 'world news breaking',
 'world news breaking headlines', 'world news cnn today',
 'world news current events', 'world news events', 'world news for this week',
 'world news for today', 'world news headline', 'world news headlines',
 'world news headlines daily nation', 'world news headlines today live',
 'world news highlights', 'world news latest headlines', 'world news now',
 'world news now cnn', 'world news recent', 'world news report',
 'world news sites', 'world news sources', 'world news stories',
 'world news today', 'world news today 2014', 'world news today cnn',
 'world news today headlines', 'world news today live',
 'world news today video', 'world news update', 'world news update today',
 'world news updates', 'world news updates today', 'world news video',
 'world news videos', 'world news website', 'world news websites',
 'world newspaper', 'world newspaper articles', 'world newspaper online',
 'world recent news', 'world times news', 'world today news', 'world top news',
 'world top news today', 'world updated news', 'world wide latest news',
 'world wide news today', 'worlds news', 'worlds news headlines',
 'worlds news today', 'worldwide breaking news', 'worldwide news today',
 'www.headline news today', 'www.headlines news', 'www.news headlines today',
 'www.news today.in', 'www.today news paper.com', 'www.todays news headlines',
 'www.todays news headlines.com', 'www.todays news.com',
 'www.world latest news']

#data = pd.read_csv('myfile.csv')['SomeColumn'].drop_duplicates().reset_index(drop=True).to_frame()[:10000]
#data.columns = ['Keyword']
#data = data['Keyword'].tolist()

stemmer = PorterStemmer()

stemmed_data = [stemmer.stem_word(word) for word in data]

hasher = HashingVectorizer(stop_words='english', ngram_range=(1,2), analyzer='word')
hash_matrix = hasher.transform(stemmed_data)

aff = cluster.AffinityPropagation()
aff.fit(hash_matrix)

df = pd.DataFrame({'Keyword': data, 'Cluster': aff.labels_.tolist()})
grouped = df.groupby('Cluster').apply(lambda frame: frame['Keyword']).reset_index(1, drop=True).to_frame('Keyword')

Answer 1

亲和传播需要二次记忆来存储完整的距离矩阵。

因此，如果您有 10000 个样本和双精度，则需要大约 800,000,000 个字节。如果在某个时候需要制作此矩阵的副本，您很容易需要 1.6 GB RAM（不包括您的输入数据和任何开销）。

如果您想达到 "hundreds of thousands"，至少要再增加 100 倍，即 80 到 160 GB RAM.