-
Notifications
You must be signed in to change notification settings - Fork 0
/
104 Predictions v2.R
443 lines (373 loc) · 11.5 KB
/
104 Predictions v2.R
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
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
if(!require(dplyr)) {
install.packages("dplyr")
library("dplyr")
}
if (!require(tidyverse)) {
install.packages("tidyverse")
library("tidyverse")
}
if (!require(caret)) {
install.packages("caret")
library("caret")
}
if (!require(doParallel)) {
install.packages("doParallel")
library("doParallel")
}
source('Definitions.R')
source('Settings.R')
source('SupportFunctions.R')
source('Colours.R')
forceRecalc <- FALSE
# Make sure we use parallel processing to speed up things
# Note that we MUST NOT set up a cluster etc - registerDoParallel() is all that
# is required. Doing anything else slows down processing!
registerDoParallel()
# Load data
load(paste0(folderTemp, 'dictKeyStats.RData'))
load(paste0(folderRData, 'MPs.RData'))
load(paste0(folderRData, 'politicians.RData'))
load(paste0(folderRData, 'votes.RData'))
load(paste0(folderRData, strDate, '-motions.RData'))
load(paste0(folderRData, 'constituencies.RData'))
mRegions <- read.csv(
file = paste0(folderTemp, '20170228_BTW17_WKr_Gemeinden_ASCII.csv'),
skip = 7,
sep = ';',
header = TRUE
)
# QUESTION Can we predict which MPs will fall into the 'frequent no show'
# category?
politicians$party.label <- sapply(politicians$party.label, standardisePartyNames)
constituencies <- constituencies %>%
left_join(mRegions, by = c('number' = 'Wahlkreis.Nr')) %>%
select(c('id', 'Landname')) %>%
unique()
inactiveMPs <- votes %>%
left_join(MPs, by=c('mandate.id' = 'id')) %>%
left_join(politicians, by=c('politician.id'='id')) %>%
filter(is.na(label)) %>%
select(c('mandate.id')) %>%
unique()
activeMPs <- MPs
nsVotes <- activeMPs %>% left_join(votes, by = c('id' = 'mandate.id')) %>%
filter(vote == 'no_show') %>%
group_by(id) %>%
dplyr::summarize(count = n(), .groups = 'keep')
activeMPs <- activeMPs %>%
left_join(nsVotes, by = c('id' = 'id'))
activeMPs[is.na(activeMPs$count), 'count'] <- 0
activeMPs <- activeMPs %>%
mutate(share = count / nrow(motions))
activeMPs <- activeMPs %>%
left_join(politicians, by = c('politician.id' = 'id'))
activeMPs <- activeMPs %>%
left_join(constituencies, by = c('electoral_data.constituency.id' = 'id'))
# Tidy up a few fields
activeMPs$Landname <- sapply(activeMPs$Landname, shortenStateNames)
activeMPs$fraction_membership.label <-
lapply(activeMPs$fraction_membership.label,
doConsolidateFraktionLabels)
# Is this MP more frequently absent?
activeMPs$isNSMP <- activeMPs$share >= dictKeyStats['cutoffMissedVotes']
# How did the MP come into their role: via parties' candidate lists, via a
# direct mandate, or as a replacement for a departing MP?
# For background on the German federal electoral system, see e.g.
# https://en.wikipedia.org/wiki/Electoral_system_of_Germany
activeMPs$listMP <- activeMPs$electoral_data.mandate_won == 'list'
activeMPs$listMP[is.na(activeMPs$listMP)] <- FALSE
activeMPs$constituencyMP <- activeMPs$electoral_data.mandate_won == 'constituency'
activeMPs$constituencyMP[is.na(activeMPs$constituencyMP)] <- FALSE
activeMPs$otherMP <- !(activeMPs$listMP | activeMPs$constituencyMP)
# Get everything set up for predicting whether an MP is a frequent no-show
YCol <- 'isNSMP'
XColsMPsConsolidated <- c(
'electoral_data.list_position',
'electoral_data.constituency_result',
'fraction_membership.label',
'sex',
'party.label',
'age',
'answeredShare',
'Landname',
'listMP',
'constituencyMP',
'otherMP'
)
XColsMPsOneHot <- c('fraction_membership.label',
'sex',
'party.label',
'Landname')
dfTemp <- activeMPs %>%
select(union(YCol, XColsMPsConsolidated)) %>% unique()
activeMPOneHot <- makeOneHot(dfTemp, XColsMPsOneHot)
XCols <- union(
setdiff(XColsMPsConsolidated, XColsMPsOneHot),
getOneHotColnames(activeMPOneHot, XColsMPsOneHot)
)
# Try a selection of common ML algorithms
lMethodSpecs1 <- list(
list(
'trainMethod' = 'glm',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'rda',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'bayesglm',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'lda',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'knn',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'cforest',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'treebag',
'sampleMethod' = 'up',
'extraParams' = NA
)
)
# This is the core sapply call where the learning is done, and performance of
# the learned models is tested. We collect results in a list that is then
# written to disk.
# The doTestMethod() function includes a number of optional parameters that were
# helpful during development. We can force a recalculation of the learning model
# by deleting the summary results file and setting the forceRecalc flag.
fName <- paste0(folderTemp, 'lConsolidatedResultsMPs.RData')
if (!file.exists(fName)) {
lConsolidatedResultsMPs <- sapply(lMethodSpecs1, function(x) {
if (!is.na(x$extraParams)) {
r <- doTestMethod(
activeMPOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0,
parse(text = x$extraParams)
)
} else {
r <- doTestMethod(
activeMPOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0
)
}
return(r[c('Training', 'Sampling', 'Precision', 'Recall')])
})
save(lConsolidatedResultsMPs, file = fName)
}
# FINDINGS None of the methods is particularly great at identifying those MPs
# that are most likely to be frequent 'no shows'
# QUESTION Does the choice of sampling method make a (positive) difference?
lMethodSpecsSamplingTests <- list(
list(
'trainMethod' = 'glm',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'glm',
'sampleMethod' = 'down',
'extraParams' = NA
),
list(
'trainMethod' = 'glm',
'sampleMethod' = 'smote',
'extraParams' = NA
),
list(
'trainMethod' = 'glm',
'sampleMethod' = 'rose',
'extraParams' = NA
)
)
fName <- paste0(folderTemp, 'lMPResultsSamplingTests.RData')
if (!file.exists(fName)) {
lMPResultsSamplingTests <-
sapply(lMethodSpecsSamplingTests, function(x) {
if (!is.na(x$extraParams)) {
r <- doTestMethod(
activeMPOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0,
parse(text = x$extraParams)
)
} else {
r <- doTestMethod(
activeMPOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0
)
}
return(r[c('Training', 'Sampling', 'Precision', 'Recall')])
})
save(lMPResultsSamplingTests, file = fName)
}
# FINDINGS Upsampling is actually best
# QUESTION Can we predict a no-show vote? First, let's try without reference to
# the isNSMP flag
# We need to set up a dataframe with all relevant predictors included
XColsVotes <- c('weekday',
'field_committees.label',
'field_topics.label')
XCols <- c(XColsMPsConsolidated, XColsVotes)
activeVotes <- activeMPs %>%
left_join(votes, by = c('id' = 'mandate.id')) %>%
left_join(motions, by = c('poll.id' = 'id')) %>%
mutate(isNSVote = (vote == 'no_show')) %>%
filter(!is.na(poll.id))
# Categorical predictors need to be one-hot encoded
XColsVotesOneHot <- c(
'fraction_membership.label',
'sex',
'party.label',
'Landname',
'weekday',
'field_committees.label',
'field_topics.label'
)
YCol <- 'isNSVote'
dfTemp <- activeVotes %>%
select(all_of(c(
YCol, XColsMPsConsolidated, XColsVotesOneHot
))) %>%
unique()
activeVotesOneHot <- makeOneHot(dfTemp, XColsVotesOneHot)
XCols <- union(
setdiff(XCols, XColsVotesOneHot),
getOneHotColnames(activeVotesOneHot, XColsVotesOneHot)
)
# We specify a (shortened) list of ML algorithms: my tests have shown that some
# of the other algorithms are too computing intensive, at least on my laptop
lMethodSpecs2 <- list(
list(
'trainMethod' = 'glm',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'lda',
'sampleMethod' = 'up',
'extraParams' = NA
),
list(
'trainMethod' = 'bayesglm',
'sampleMethod' = 'up',
'extraParams' = NA
)
)
fName <- paste0(folderTemp, 'lResultsVotes.RData')
if (!file.exists(fName)) {
lResultsVotes <- sapply(lMethodSpecs2, function(x) {
if (!is.na(x$extraParams)) {
r <- doTestMethod(
activeVotesOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0,
parse(text = x$extraParams)
)
} else {
r <- doTestMethod(
activeVotesOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 0
)
}
return(r[c('Training', 'Sampling', 'Precision', 'Recall')])
})
save(lResultsVotes, file = fName)
}
# FINDINGS Overall, these are not great results in predicting no-show behaviour
# We hypothesise that a number of factors outside the dataset used here likely
# play a significant role: this would need to be analysed further in subsequent
# work (out of scope of this capstone project)
# QUESTION Does adding a flag that highlights the most conspicuously no-show MPs
# improve predictive performance?
XCols <- append(XCols, 'isNSMP')
dfTemp <- activeVotes %>%
select(all_of(c(
YCol, XColsMPsConsolidated, 'isNSMP', XColsVotesOneHot
))) %>%
unique()
activeVotesOneHot <- makeOneHot(dfTemp, XColsVotesOneHot)
fName <- paste0(folderTemp, 'lResultsVotesEnhanced.RData')
if (!file.exists(fName)) {
lResultsVotesEnhanced <- sapply(lMethodSpecs2, function(x) {
print(paste0('Running ', x$trainMethod, '(', x$sampleMethod, ')'))
if (!is.na(x$extraParams)) {
r <- doTestMethod(
activeVotesOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 1,
parse(text = x$extraParams)
)
} else {
r <- doTestMethod(
activeVotesOneHot,
XCols,
YCol,
trainMethodName = x$trainMethod,
samplingMethodName = x$sampleMethod,
forceCalc = get0('forceRecalc', ifnotfound = FALSE),
returnModel = FALSE,
versionTag = 1
)
}
return(r[c('Training', 'Sampling', 'Precision', 'Recall')])
})
save(lResultsVotesEnhanced, file = fName)
}
# FINDINGS Sadly, the flag does not make a meaningful difference
# We now stop the (implicitly created) local multi-processor cluster
stopImplicitCluster()
# ... and we are done!
print('Done.\n\n')