/
gamboost.R
202 lines (153 loc) · 7.3 KB
/
gamboost.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
#############
# Use Mboost..... and caret?
# Still a work in progress...
#############
library(geoGAM)
library(mboost)
#Read and format data
data.gam <- read.csv("Datafiles/data.national.6.5.csv")
data.gam$lith.pred <- factor(data.gam$lith.pred)
data.gam$Primary.rocktype <- factor(data.gam$Primary.rocktype)
data.gam$Secondary.rocktype <- factor(data.gam$Secondary.rocktype)
data.gam$usda_tex <- factor(data.gam$usda_tex)
#PA$pres <- factor(PA$pres,levels=c("Present",'Absent'))
data.gam <- data.gam[!is.na(data.gam$IV),]
data.gam.PA <- data.gam[,!colnames(data.gam) %in% c("IV","PlotID","predictedClass","Present","Absent","asp_val")] #remove unneeded variables
#data.gam.PA$pres <- ifelse(data.gam.PA$pres=="Present",1,0)
#data.gam.PA <- data.gam.PA[,!colnames(data.gam.PA) %in% c("pres","PlotID","predictedClass","Present","Absent","asp_val")] #remove unneeded variables
data.gam.PA <- data.gam.PA[complete.cases(data.gam.PA),c(6,1:5,8:26,28,31,32,34,7,27,29,30,33)] #reorder to put response first, then spatial vars, then continuous then factors
data.gam.PA <- data.gam.PA[data.gam.PA$usda_tex != 13,]
colnames(data.gam.PA)[c(2,3)]<- c("x","y")
data.gam.PA$usda_tex <- factor(data.gam.PA$usda_tex)
#data.gam <- data.gam[,!colnames(data.gam) %in% c("pres","PlotID","predictedClass","Present","Absent","asp_val")] #remove unneeded variables
#data.gam <- data.gam[complete.cases(data.gam),c(7,1:5,8:26,28,31,32,34,6,27,29,30,33)] #reorder to put response first, then spatial vars, then continuous then factors
#data.gam <- data.gam[data.gam$usda_tex != 13,]
#colnames(data.gam)[c(2,3)]<- c("x","y")
#data.gam$usda_tex <- factor(data.gam$usda_tex)
#add two-way interactions of continuous variables
#data.gam_cont <- data.gam[,7:29]
#data.gam_inter <- model.matrix(~.^2-1,data.gam_cont)
#colnames(data.gam_inter) <- gsub(":","x",colnames(data.gam_inter))
#data.gam2 <- cbind(data.gam[,c(1:6,30:ncol(data.gam))],data.gam_inter)
#data.gam2 <- cbind(data.gam[,c(1:6,30:ncol(data.gam))],data.gam_inter[,1:30]) #for testing purposes
#Split 80/20 into training and test sets
set.seed(13537)
intrain <- createDataPartition(y=data.gam.PA$pres,p=0.8,list=FALSE)
data.gam.PA.train <- data.gam.PA[intrain,]
#data.train.c <- data.train[complete.cases(data.train),]
#data.train.nonzero <- subset(data.train,IV>0)
data.gam.PA.test <- data.gam.PA[-intrain,]
#Split 80/20 into training and test sets
#set.seed(13537)
#intrain <- createDataPartition(y=data.gam.PA2$IV,p=0.8,list=FALSE)
#data.gam.PA2.train <- data.gam.PA2[intrain,]
#data.train.c <- data.train[complete.cases(data.train),]
#data.train.nonzero <- subset(data.train,IV>0)
#data.gam.PA2.test <- data.gam.PA2[-intrain,]
#### Create formula
makeMboostForm <- function(df){
resp <- paste(colnames(df)[1],"~")
names <- colnames(df)
coords <- c("x","y")
l.fact <- names(df)[ unlist( lapply( df, is.factor) ) == TRUE ]
factors <- grep( paste0("^", resp, "$"), l.fact, invert = T, value = T)
factors <- factors[-which(factors=="pres")]
print(factors)
nums <- names[-which(names %in% c(colnames(df)[1],coords,factors))]
#Make splines
form.bbs <- paste( "bbs(",
paste(nums,
collapse = paste(", center = T, df = 1)", " + bbs("
)),
", center = T, df = 1)",
sep = ""
)
#Make spatial
form.spatial <- "bspatial(x,y,df=6,knots=12)"
#Make nonstationary
form.nonspatial <- paste( "bspatial(", paste(coords, collapse = ","), ", by=",
paste(nums,
collapse = paste(
", df=6, knots = 12) ", " + bspatial(", paste(coords, collapse = ","), ", by=" )
),
", df=6, knots = 12)",
sep = ""
)
#make factors - and linear continuous base learners
form.bols <- paste( "bols(",
paste(c(factors,nums),
collapse = paste(", intercept = T, df = 1)", " + bols("
)),
", intercept = T, df = 1)",
sep = "")
return(paste(resp,paste(form.bbs,form.spatial,form.bols,form.nonspatial,sep="+")))
}
form <- makeMboostForm(data.gam.PA.train)
form
form <- as.formula(form)
#form = "pres ~ bbs(STDAGE,center=T,df=1)+bspatial(x,y,df=1)"
#form[[2]]
m1 <- gamboost(formula=as.formula(makeMboostForm(data.gam.PA.train)),data=data.gam.PA.train,family=Binomial(),control=boost_control(mstop=700))
c1 <- makeCluster(round(detectCores()*.7))
registerDoParallel(c1)
cv <- cvrisk(m1)
stopCluster(c1)
registerDoParallel()
plot(cv)
finalModel <- m1[mstop(cv)]
summary(finalModel)
#model <- gamboost()
#Test set performance
predicted.PA.test <- predict(finalModel,newdata=data.gam.PA.test,type="response",na.action=na.pass) #predict on test set
roc.obj <- roc(ifelse(data.gam.PA.test$pres == "Present",1,0),pred=predicted.PA.test)
auc(roc.obj)
predicted.PA.train <- predict(finalModel,newdata=data.gam.PA.train,type="response",na.action=na.pass) #predict on test set
roc.obj <- roc(ifelse(data.gam.PA.train$pres == "Present",1,0),pred=predicted.PA.train)
auc(roc.obj)
############
#basic logistic regression w/smote- for ROC AUC comparison
ctrlPar <- trainControl(method='repeatedcv',number=5, repeats=5,allowParallel=TRUE,classProbs=TRUE,summaryFunction=twoClassSummary,sampling="smote")
library(doParallel)
c1 <- makeCluster(round(detectCores()*.5))
registerDoParallel(c1)
set.seed(8081)
logistic <- train(data.gam.PA.train[,-1],data.gam.PA.train[,1],method="glm",family="binomial",
trControl=trainControl(method="none"))
#logistic <- glm(pres~.,data=data.gam.PA.train,family=binomial)
stopCluster(c1)
registerDoParallel()
predicted.PA.test <- predict(logistic,newdata=data.gam.PA.test,type="prob",na.action=na.pass) #predict on test set
roc.obj <- roc(ifelse(data.gam.PA.test$pres == "Present",1,0),pred=predicted.PA.test,metric="ROC")
auc(roc.obj) #AUC = 0.8922
summary(logistic)
#GAM (not boosted)
library(caret)
c1 <- makeCluster(round(detectCores()*.5))
registerDoParallel(c1)
set.seed(8081)
gam <- train(data.gam.PA.train[,-1],data.gam.PA.train[,1],method="gam",family=binomial,
trControl=ctrlPar,tuneGrid=expand.grid(select=T,method="GCV.Cp"),metric="ROC")
stopCluster(c1)
registerDoParallel()
###random Forest w/smote - CV AUC = 0.9114
rfGrid <- expand.grid(mtry=c(1:3))
set.seed(8081)
c1 <- makeCluster(round(detectCores()*.5))
registerDoParallel(c1)
rfTune <- train(x=data.gam.PA.train[,-1],y=data.gam.PA.train[,1],method='rf', trControl=ctrlPar,metric="ROC",ntree=200,tuneGrid=rfGrid)
rfTune
ggplot(rfTune)
plot(rfTune$finalModel)
stopCluster(c1)
registerDoParallel()
#test set accuracy
predicted.PA.test <- predict(rfTune,newdata=data.gam.PA.test,type="prob",na.action=na.pass) #predict on test set
roc.obj <- roc(ifelse(data.gam.PA.test$pres == "Present",1,0),pred=predicted.PA.test$Present,metric="ROC")
auc(roc.obj) #AUC = 0.9182
varImp(rfTune)
library(randomForest)
partialPlot(rfTune$finalModel, data.gam.PA.train[,-1], x.var = "bio1",which.class="Present")
partialPlot(rfTune$finalModel, data.gam.PA.train[,-1], x.var = "bio5",which.class="Present")
partialPlot(rfTune$finalModel, data.gam.PA.train[,-1], x.var = "bio11",which.class="Present")
x <- partialPlot(rfTune$finalModel, data.gam.PA.train[,-1], x.var = "bio5",which.class="Present",plot=F)
plot(x$x,inv.logit(x$y),type='l')