forked from justmarkham/DAT8
-
Notifications
You must be signed in to change notification settings - Fork 0
/
16_kaggle.py
338 lines (245 loc) · 10.6 KB
/
16_kaggle.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
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
'''
CLASS: Kaggle Stack Overflow competition
'''
# read in the file and set the first column as the index
import pandas as pd
train = pd.read_csv('train.csv', index_col=0)
train.head()
'''
What are some assumptions and theories to test?
OwnerUserId: not unique within the dataset, assigned in order
OwnerCreationDate: users with older accounts have more open questions
ReputationAtPostCreation: higher reputation users have more open questions
OwnerUndeletedAnswerCountAtPostTime: users with more answers have more open questions
Title and BodyMarkdown: well-written questions are more likely to be open
Tags: 1 to 5 tags are required, many unique tags
OpenStatus: most questions should be open (encoded as 1)
'''
## OPEN STATUS
# dataset is perfectly balanced in terms of OpenStatus (not a representative sample)
train.OpenStatus.value_counts()
## USER ID
# OwnerUserId is not unique within the dataset, let's examine the top user
train.OwnerUserId.value_counts()
# mostly closed questions, few answers, all lowercase, grammatical mistakes
train[train.OwnerUserId==466534].describe()
train[train.OwnerUserId==466534].head(10)
# let's find a user with a high proportion of open questions
train.groupby('OwnerUserId').OpenStatus.mean()
train.groupby('OwnerUserId').OpenStatus.agg(['mean','count']).sort('count')
# lots of answers, better grammar, multiple tags, all .net
train[train.OwnerUserId==185593].head(10)
## REPUTATION
# ReputationAtPostCreation is higher for open questions: possibly use as a feature
train.groupby('OpenStatus').ReputationAtPostCreation.describe().unstack()
# not a useful histogram
train.ReputationAtPostCreation.plot(kind='hist')
# much more useful histogram
train[train.ReputationAtPostCreation < 1000].ReputationAtPostCreation.plot(kind='hist')
# grouped histogram
train[train.ReputationAtPostCreation < 1000].hist(column='ReputationAtPostCreation', by='OpenStatus', sharey=True)
# grouped box plot
train[train.ReputationAtPostCreation < 1000].boxplot(column='ReputationAtPostCreation', by='OpenStatus')
## ANSWER COUNT
# rename column
train.rename(columns={'OwnerUndeletedAnswerCountAtPostTime':'Answers'}, inplace=True)
# Answers is higher for open questions: possibly use as a feature
train.groupby('OpenStatus').Answers.describe().unstack()
## TITLE LENGTH
# create a new feature that represents the length of the title (in characters)
train['TitleLength'] = train.Title.apply(len)
# Title is longer for open questions: possibly use as a feature
train.groupby('OpenStatus').TitleLength.describe().unstack()
train.boxplot(column='TitleLength', by='OpenStatus')
'''
Define a function that takes a raw CSV file and returns a DataFrame that
includes all created features (and any other modifications)
'''
# define the function
def make_features(filename):
df = pd.read_csv(filename, index_col=0)
df.rename(columns={'OwnerUndeletedAnswerCountAtPostTime':'Answers'}, inplace=True)
df['TitleLength'] = df.Title.apply(len)
return df
# apply function to both training and testing files
train = make_features('train.csv')
test = make_features('test.csv')
'''
Evaluate a model with three features
'''
# define X and y
feature_cols = ['ReputationAtPostCreation', 'Answers', 'TitleLength']
X = train[feature_cols]
y = train.OpenStatus
# split into training and testing sets
from sklearn.cross_validation import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=1)
# fit a logistic regression model
from sklearn.linear_model import LogisticRegression
logreg = LogisticRegression(C=1e9)
logreg.fit(X_train, y_train)
# examine the coefficients to check that they makes sense
logreg.coef_
# predict response classes and predict class probabilities
y_pred_class = logreg.predict(X_test)
y_pred_prob = logreg.predict_proba(X_test)[:, 1]
# check how well we did
from sklearn import metrics
metrics.accuracy_score(y_test, y_pred_class) # 0.564 (better than guessing)
metrics.confusion_matrix(y_test, y_pred_class) # predicts closed a lot of the time
metrics.roc_auc_score(y_test, y_pred_prob) # 0.591 (not horrible)
metrics.log_loss(y_test, y_pred_prob) # 0.684 (what is this?)
# let's see if cross-validation gives us similar results
from sklearn.cross_validation import cross_val_score
scores = cross_val_score(logreg, X, y, scoring='log_loss', cv=10)
scores.mean() # 0.684 (identical to train/test split)
scores.std() # very small
'''
Understanding log loss
'''
# 5 pretend response values
y_test = [0, 0, 0, 1, 1]
# 5 sets of predicted probabilities for those observations
y_pred_prob_sets = [[0.1, 0.2, 0.3, 0.8, 0.9],
[0.4, 0.4, 0.4, 0.6, 0.6],
[0.4, 0.4, 0.7, 0.6, 0.6],
[0.4, 0.4, 0.9, 0.6, 0.6],
[0.5, 0.5, 0.5, 0.5, 0.5]]
# calculate AUC for each set of predicted probabilities
for y_pred_prob in y_pred_prob_sets:
print y_pred_prob, metrics.roc_auc_score(y_test, y_pred_prob)
# calculate log loss for each set of predicted probabilities
for y_pred_prob in y_pred_prob_sets:
print y_pred_prob, metrics.log_loss(y_test, y_pred_prob)
'''
Create a submission file
'''
# train the model on ALL data (not X_train and y_train)
logreg.fit(X, y)
# predict class probabilities for the actual testing data (not X_test)
X_oos = test[feature_cols]
oos_pred_prob = logreg.predict_proba(X_oos)[:, 1]
# sample submission file indicates we need two columns: PostId and predicted probability
test.index # PostId
oos_pred_prob # predicted probability
# create a DataFrame that has 'id' as the index, then export to a CSV file
sub = pd.DataFrame({'id':test.index, 'OpenStatus':oos_pred_prob}).set_index('id')
sub.to_csv('sub1.csv') # 0.687
'''
Explore data and create more features
'''
## TAGS
# Tag1 is required, and the rest are optional
train.isnull().sum()
# create a new feature that represents the number of tags for each question
train['NumTags'] = train.loc[:, 'Tag1':'Tag5'].notnull().sum(axis=1)
# NumTags is higher for open questions: possibly use as a feature
train.groupby('OpenStatus').NumTags.mean()
train.groupby('NumTags').OpenStatus.mean()
## USER ID
# OwnerUserId is assigned in numerical order
train.sort('OwnerUserId').OwnerCreationDate
# OwnerUserId is lower for open questions: possibly use as a feature
train.groupby('OpenStatus').OwnerUserId.mean()
# account age at time of question is probably a better feature
train['OwnerCreationDate'] = pd.to_datetime(train.OwnerCreationDate)
train['PostCreationDate'] = pd.to_datetime(train.PostCreationDate)
train['OwnerAge'] = (train.PostCreationDate - train.OwnerCreationDate).dt.days
# check that it worked
train.head()
train.boxplot(column='OwnerAge', by='OpenStatus')
train[train.OwnerAge < 0].head()
# fix errors
import numpy as np
train['OwnerAge'] = np.where(train.OwnerAge < 0, 0, train.OwnerAge)
train.boxplot(column='OwnerAge', by='OpenStatus')
'''
Evaluate new set of features using cross-validation
'''
feature_cols = ['ReputationAtPostCreation', 'Answers', 'TitleLength', 'NumTags', 'OwnerAge']
X = train[feature_cols]
cross_val_score(logreg, X, y, scoring='log_loss', cv=10).mean() # 0.675
'''
Update make_features and create another submission file
'''
# update the function
def make_features(filename):
df = pd.read_csv(filename, index_col=0, parse_dates=['OwnerCreationDate', 'PostCreationDate'])
df.rename(columns={'OwnerUndeletedAnswerCountAtPostTime':'Answers'}, inplace=True)
df['TitleLength'] = df.Title.apply(len)
df['NumTags'] = df.loc[:, 'Tag1':'Tag5'].notnull().sum(axis=1)
df['OwnerAge'] = (df.PostCreationDate - df.OwnerCreationDate).dt.days
df['OwnerAge'] = np.where(df.OwnerAge < 0, 0, df.OwnerAge)
return df
# apply function to both training and testing files
train = make_features('train.csv')
test = make_features('test.csv')
# train the model on ALL data
feature_cols = ['ReputationAtPostCreation', 'Answers', 'TitleLength', 'NumTags', 'OwnerAge']
X = train[feature_cols]
logreg.fit(X, y)
# predict class probabilities for the actual testing data
X_oos = test[feature_cols]
oos_pred_prob = logreg.predict_proba(X_oos)[:, 1]
# create submission file
sub = pd.DataFrame({'id':test.index, 'OpenStatus':oos_pred_prob}).set_index('id')
sub.to_csv('sub2.csv') # 0.650
'''
Build a document-term matrix from Title using CountVectorizer
'''
# use CountVectorizer with the default settings
from sklearn.feature_extraction.text import CountVectorizer
vect = CountVectorizer()
dtm = vect.fit_transform(train.Title)
# define X and y
X = dtm
y = train.OpenStatus
# slightly improper cross-validation of a Naive Bayes model
from sklearn.naive_bayes import MultinomialNB
nb = MultinomialNB()
cross_val_score(nb, X, y, scoring='log_loss', cv=10).mean() # 0.657
# try tuning CountVectorizer and repeat Naive Bayes
vect = CountVectorizer(stop_words='english')
dtm = vect.fit_transform(train.Title)
X = dtm
cross_val_score(nb, X, y, scoring='log_loss', cv=10).mean() # 0.635
# build document-term matrix for the actual testing data and make predictions
nb.fit(X, y)
oos_dtm = vect.transform(test.Title)
oos_pred_prob = nb.predict_proba(oos_dtm)[:, 1]
sub = pd.DataFrame({'id':test.index, 'OpenStatus':oos_pred_prob}).set_index('id')
sub.to_csv('sub3.csv') # 0.544
'''
BONUS: Dummy encoding of Tag1
'''
# number of unique tags for Tag1 (over 5000)
train.Tag1.nunique()
# percentage of open questions varies widely by tag
train.groupby('Tag1').OpenStatus.agg(['mean','count']).sort('count')
# convert Tag1 from strings to integers
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
train['Tag1_enc'] = le.fit_transform(train.Tag1)
# confirm that the conversion worked
train.Tag1.value_counts().head()
train.Tag1_enc.value_counts().head()
# create a dummy column for each value of Tag1_enc (returns a sparse matrix)
from sklearn.preprocessing import OneHotEncoder
ohe = OneHotEncoder()
tag1_dummies = ohe.fit_transform(train[['Tag1_enc']])
tag1_dummies
# define X and y
X = tag1_dummies
y = train.OpenStatus
# try a Naive Bayes model with tag1_dummies as the features
cross_val_score(nb, X, y, scoring='log_loss', cv=10).mean() # 0.650
# adjust Tag1 on testing set since LabelEncoder errors on new values during a transform
test['Tag1'] = test['Tag1'].map(lambda s: '<unknown>' if s not in le.classes_ else s)
le.classes_ = np.append(le.classes_, '<unknown>')
# apply the same encoding to the actual testing data and make predictions
nb.fit(X, y)
test['Tag1_enc'] = le.transform(test.Tag1)
oos_tag1_dummies = ohe.transform(test[['Tag1_enc']])
oos_pred_prob = nb.predict_proba(oos_tag1_dummies)[:, 1]
sub = pd.DataFrame({'id':test.index, 'OpenStatus':oos_pred_prob}).set_index('id')
sub.to_csv('sub4.csv') # 0.652