appliedpredictivemodeling/ch11.R at master · BenTremblay1/appliedpredictivemodeling · GitHub

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
# CH 11

library("AppliedPredictiveModeling")
library("caret")
library("klaR")
library("MASS")
library("pROC")
library("randomForest")

# generate simulated data

set.seed(975)
simulatedTrain <- quadBoundaryFunc(500)
simulatedTest  <- quadBoundaryFunc(1000)

# fit random forest and quadradic discriminant model

rfModel <- randomForest(class ~ X1 + X2,
                        data = simulatedTrain,
                        ntree = 2000)

qdaModel <- qda(class ~ X1 + X2,
                data = simulatedTrain)

qdaTrainPred <- predict(qdaModel, simulatedTrain)
str(qdaTrainPred)
head(qdaTrainPred$posterior)

qdaTestPred <- predict(qdaModel, simulatedTest)
simulatedTrain$QDAprob <- qdaTrainPred$posterior[, 'Class1']
simulatedTest$QDAprob  <- qdaTestPred$posterior[, 'Class1']

# return probabilities of each class
rfTestPred <- predict(rfModel, simulatedTest, type='prob')
head(rfTestPred)
simulatedTest$RFprob <- rfTestPred[, 'Class1']

# return predicted class
simulatedTest$RFclass <- predict(rfModel, simulatedTest)
head(simulatedTest$RFclass)

# caret can calculate sensitivity and specificity!
sensitivity(data = simulatedTest$RFclass,
            reference = simulatedTest$class,
            positive = 'Class1')

specificity(data = simulatedTest$RFclass,
            reference = simulatedTest$class,
            negative = 'Class2')

posPredValue(data = simulatedTest$RFclass,
            reference = simulatedTest$class,
            positive = 'Class1')

negPredValue(data = simulatedTest$RFclass,
            reference = simulatedTest$class,
            positive = 'Class2')

# caret can do confusion matrix too!

# most analyses should just start with this function. It provides the PPV, NPV
# and an assortment of other good data to know if the model is performing to
# the goals of the analysis

confusionMatrix(data = simulatedTest$RFclass,
                reference = simulatedTest$class,
                positive = 'Class1')


# generating ROC curves

# create plottable object
rocCurve <- roc(response = simulatedTest$class,
                predictor = simulatedTest$RFprob,
                # reverse labels
                levels = rev(levels(simulatedTest$class)))
head(rocCurve)
str(rocCurve)

auc(rocCurve)
ci.roc(rocCurve)

plot(rocCurve, legacy.axes = TRUE)

# lift charts:

labs <- c(RFprob = 'Random Forest',
          QDAprob = 'Quadradic Disciminant Analysis')

liftCurve <- lift(class ~ RFprob + QDAprob,
                  data = simulatedTest,
                  labels = labs)
liftCurve
xyplot(liftCurve,
       auto.key = list(columns = 2,
                       lines = TRUE,
                       points = FALSE))

# calibration plots

# build the object
calCurve <- calibration(class ~ RFprob + QDAprob, data = simulatedTest)
calCurve

xyplot(calCurve, auto.key = list(columns = 2))

# try fitting a sigmoid plot using the glm package

sigmoidalCal <- glm(relevel(class, ref = "Class2") ~ QDAprob,
                    data = simulatedTrain,
                    family = binomial)
coef(summary(sigmoidalCal))

sigmoidProbs <- predict(sigmoidalCal,
                        newdata = simulatedTest[, "QDAprob", drop = FALSE],
                        type = "response")
simulatedTest$QDAsigmoid <- sigmoidProbs

BayesCal <- NaiveBayes(class ~ QDAprob, data = simulatedTrain,
                       usekernel = TRUE)
BayesProbs <- predict(BayesCal,
                      newdata = simulatedTest[, 'QDAprob', drop = FALSE])
simulatedTest$QDABayes <- BayesProbs$posterior[, 'Class1']
head(BayesProbs$posterior)
head(simulatedTest[, c(5:6, 8, 9)])

calCurve2 <- calibration(class ~ QDAprob + QDABayes + QDAsigmoid,
                         data = simulatedTest)
xyplot(calCurve2)

# interesting ... play with the lift curve a little more
labs <- c(QDABayes = 'Bayes - QDA',
          QDAsigmoid= 'Quadradic Disciminant Analysis w/ sigmoid',
          RFprob = 'Random Forest',
          QDAprob = 'Quadradic Disciminant Analysis')

liftCurve <- lift(class ~ QDABayes + QDAsigmoid + RFprob + QDAprob,
                  data = simulatedTest,
                  labels = labs)
liftCurve
xyplot(liftCurve,
       auto.key = list(columns = 2,
                       lines = TRUE,
                       points = FALSE))


# play around with the ROC curve with the various models created:

rocCurve.rfprob <- roc(response = simulatedTest$class,
                predictor = simulatedTest$RFprob,
                # reverse labels
                levels = rev(levels(simulatedTest$class)))


rocCurve.qdaprob <- roc(response = simulatedTest$class,
                predictor = simulatedTest$QDAprob,
                # reverse labels
                levels = rev(levels(simulatedTest$class)))

rocCurve.qdabayes <- roc(response = simulatedTest$class,
                predictor = simulatedTest$QDABayes,
                # reverse labels
                levels = rev(levels(simulatedTest$class)))

rocCurve.qdasigmoid <- roc(response = simulatedTest$class,
                predictor = simulatedTest$QDAsigmoid,
                # reverse labels
                levels = rev(levels(simulatedTest$class)))

plot(rocCurve.rfprob, legacy.axes = TRUE) # from the book
plot.roc(rocCurve.qdaprob, legacy.axes = TRUE, add = TRUE,
         col = 'red')
plot.roc(rocCurve.qdabayes, legacy.axes = TRUE, add = TRUE,
         col = 'blue')
plot.roc(rocCurve.qdasigmoid, legacy.axes = TRUE, add = TRUE,
         col = 'green')