-
Notifications
You must be signed in to change notification settings - Fork 5
/
infoshieldcoarse.py
239 lines (184 loc) · 8.97 KB
/
infoshieldcoarse.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
#!/usr/bin/env python3
# Author: Catalina Vajiac
# Purpose: Coarse clustering of text documents
# Usage: ./infoshieldcoarse.py [filename]
import heapq
import math
import networkx as nx
import numpy as np
import os, sys
import pandas
import pickle
import random
import re
import scipy
import time
import matplotlib.pyplot as plt
from collections import Counter, defaultdict
from datetime import datetime
from networkx.algorithms import bipartite
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.cluster import adjusted_rand_score, homogeneity_score
# Utilities
def term_frequency(phrase, document):
''' return tf of phrase in document '''
return sum([p == phrase for p in document])
def filter_text(text):
if type(text) is not str: # nan
return ''
replace = [(r'\d+', ''), (r'[^\x00-\x7F]+', ' '), (re.compile('<.*?>'), '')]
nbsp_variants = [(' ', ''), ('nbsp;', '')]
br_variants = [('<b', ''), ('<br', ''), ('br>', '')]
for source, target in replace + nbsp_variants + br_variants:
text = re.sub(source, target, text)
return text
class InfoShieldCoarse():
def __init__(self, filename: str, doc_id_header='', doc_text_header='', num_phrases=10):
# init basic variables
self.time = time.time()
self.num_phrases = num_phrases
self.filename_full = filename.split('.')[0]
self.filename = os.path.basename(filename).split('.')[0]
#self.time_filename = '{}_streaming-{}_time.txt'.format(self.filename)
self.ngrams = (5, 5)
self.index_to_docid = Counter()
self.docid_to_index = Counter()
self.data = pandas.read_csv(filename, lineterminator='\n')
self.determine_header_names(doc_text_header, doc_id_header)
#self.data = self.data.drop_duplicates(subset=['title', self.description])
if 'timestamp' in self.data.columns:
self.data.sort_values(by=['timestamp']) # since ads not in order as they should be
self.num_ads = len(self.data.index)
self.cluster_graph = nx.Graph()
# setup tfidf
tfidf = TfidfVectorizer(token_pattern=r'[^\s]+', lowercase=False, ngram_range=self.ngrams,
sublinear_tf=True)
self.tokenizer = tfidf.build_analyzer()
self.document_freq = defaultdict(float)
self.term_freq = defaultdict(lambda: Counter())
self.length = Counter()
self.data[self.description] = self.data[self.description].apply(filter_text)
self.tfidfs = tfidf.fit_transform(self.data[self.description].values)
self.tfidf_indices = tfidf.get_feature_names_out()
self.num_ads_features = 0
def determine_header_names(self, doc_text_header: str, doc_id_header: str):
''' automatically determine relevant header names for doc id, doc text'''
columns = set(self.data.columns)
indices = {'ad_id', 'index', 'TweetID', 'id'}
descriptions = {'u_Description', 'description', 'body', 'Tweet', 'text'}
phones = {'u_PhoneNumbers', 'phone', 'PhoneNumber'}
descriptions.add(doc_text_header)
indices.add(doc_id_header)
indices.add(doc_text_header)
for name, field in [('text', descriptions), ('unique id', indices)]:#, ('phone #', phones)]:
if not len(columns.intersection(field)):
print('Add "{}" header to possible descriptions!'.format(name))
exit(1)
self.description = columns.intersection(descriptions).pop()
self.id = columns.intersection(indices).pop()
self.phone = columns.intersection(phones).pop() if len(columns.intersection(phones)) else ''
def tokenize_text(self, text):
#include_field = lambda x: x in self.data.columns and type(row[x]) == str
return self.tokenizer(filter_text(text))
def top_tfidf_phrases(self, doc_id, index: int, return_all=False):
''' return the top phrases with highest tfidf score '''
#def score(doc_id: str, phrase: str) -> float:
# return self.term_freq[doc_id][phrase] * self.length[doc_id] / self.document_freq[phrase]
#tfidf_pairs = [(score(doc_id, phrase), phrase) for phrase in phrases]
_, cols = self.tfidfs[index].nonzero()
tfidf_pairs = [(self.tfidfs[index, c], self.tfidf_indices[c]) for c in cols]
return heapq.nlargest(self.num_phrases, tfidf_pairs)
def process_ad(self, index: int, row):
''' find top phrases and add the ad to the cluster graph '''
doc_id = row[self.id]
self.index_to_docid[index] = doc_id
self.docid_to_index[doc_id] = index
if 'title' in row:
text = row['title'] if type(row['title']) == str else ''
else:
text = ''
text += ' ' + row[self.description] if type(row[self.description]) == str else ''
phrases = self.tokenize_text(text)
#top_tfidf = [phrase for _, phrase in self.top_tfidf_phrases(doc_id, set(phrases))]
top_tfidf = [phrase for _, phrase in self.top_tfidf_phrases(doc_id, index)]
self.cluster_graph.add_nodes_from(top_tfidf, bipartite=0)
self.cluster_graph.add_node(doc_id, bipartite=1)
self.cluster_graph.add_edges_from([(doc_id, phrase) for phrase in top_tfidf])
def generate_labels(self):
document_nodes = set([n for n, d in self.cluster_graph.nodes(data=True) if d['bipartite']])
self.labels = [-1]*len(self.data.index)
for i, component in enumerate(nx.connected_components(self.cluster_graph)):
docs = [c for c in component if c in document_nodes]
if len(docs) == 1:
continue
for docid in docs:
self.labels[self.docid_to_index[docid]] = i
def write_cluster_graph(self):
''' write cluster graph as pkl file '''
if not os.path.isdir('pkl_files'):
os.mkdir('pkl_files')
#with open('pkl_files/{}_ad_graph.pkl'.format(self.filename), 'wb') as f:
# pickle.dump(self.cluster_graph, f)
def write_csv_labels(self):
''' write new csv, with LSH labels '''
filename_stub = self.filename_full
self.final_data_filename = filename_stub + '_LSH_labels.csv'
self.unfiltered_data_filename = filename_stub + '_full_LSH_labels.csv'
self.data['LSH label'] = self.labels
data_filtered = self.data.dropna(subset=[self.description])
data_filtered.to_csv(self.unfiltered_data_filename, index=False)
data_filtered = data_filtered[data_filtered['LSH label'] != -1]
data_filtered.to_csv(self.final_data_filename, index=False)
def clustering(self):
''' process each ad individually and incrementally save cluster graph. '''
t = time.time()
index = 0
for _, row in self.data.iterrows():
if index and not index % 10000:
time_elapsed = time.time() - t
print(index, '/', self.num_ads, 'time', time_elapsed)
self.docid_to_index[row[self.id]] = index
self.index_to_docid[index] = row[self.id]
self.process_ad(index, row)
index += 1
self.generate_labels()
self.write_cluster_graph()
self.write_csv_labels()
self.total_time = time.time() - t
print('Finished clustering!', self.total_time)
#print(self.num_ads_features, len(self.data.index))
def get_clusters(self):
''' given cluster graph, return the relevant connected components '''
criteria = lambda x: len(x) >= 5
return [c for c in nx.connected_components(self.cluster_graph) if criteria(c)]
def get_docs(self, cluster_nodes):
''' given a set of cluster nodes, return the documents they represent '''
return cluster_nodes
def print_clusters(self):
print('number of clusters', len(self.get_clusters()))
clusters = sorted(self.get_clusters(), key=lambda x: len(self.get_docs(x)), reverse=True)
document_nodes = [n for n, d in self.cluster_graph.nodes(data=True) if d['bipartite']]
for i, cluster in enumerate(clusters):
docs = [c for c in cluster if c in document_nodes]
print('cluster:', i, 'len:', len(docs))
for doc_id in docs:
index = self.docid_to_index[doc_id]
row = self.data.loc[index]
try:
description = row[self.description]
except:
print('issue with doc_id', doc_id, 'and desc', self.description)
print(doc_id, [n for n in self.cluster_graph.neighbors(doc_id)], row['label'])
#print(doc_id, [n for n in self.cluster_graph.neighbors(doc_id)], row['is_spam'])
print()
print('\n\n')
def usage(exit_code):
print('Usage: _ [filename]')
exit(exit_code)
if __name__ == '__main__':
# either just provide filename, or provide all params
if len(sys.argv) not in [2, 5]:
usage(1)
filename = sys.argv[1]
d = InfoShieldCoarse(filename, num_phrases=10)
d.clustering()