Coursera_PracticalMachineLearning/WriteUp.Rhtml at master · jbregli/Coursera_PracticalMachineLearning · 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
<html>

<head>
<title>Practical Machine Learning Assignment</title>
</head>

<body>

<p>This document sums up what I have done on the programming assignment for the course "Practical Machine Learning" on Coursera</p>

<h1>Step 0:</h1>

<p>Let's first import the library we will use. </p>
<!--begin.rcode
library(lattice); library(ggplot2); library(caret);library(kernlab); library(randomForest)
end.rcode-->

<p>And let's fix the seed to make the experience reproductive. </p>
<!--begin.rcode
set.seed(32323)
end.rcode-->

<h1>Step 1: Loading the data</h1>

<p>Let's load the train and test datasets.</p>

<!--begin.rcode
# Load the training dataset:
dataset  <- read.csv("pml-training.csv", na.strings=c("NA",""), strip.white=TRUE)
dim(dataset)

# Load the test dataset
dataset_test <- read.csv("pml-testing.csv", na.strings=c("NA",""), strip.white=T)
dim(dataset_test)
end.rcode-->

<p>We have <b>19622 training examples</b> composed of the recording of <b>160 measures</b>. <br/>
We have <b>20 test examples</b>.</p>

<h1>Step 2: Cleaning the data</h1>

<p>By looking at the data we can notice that there are a lot of empty columns.<br/>
Let's remove them. </p>

<!--begin.rcode
# Cleaning the data:
isNA <- apply(dataset, 2, function(x) { sum(is.na(x)) })

# For the training dataset:
dataset <- subset(dataset[, which(isNA == 0)],
                    select=-c(X, user_name, new_window, num_window,
                              raw_timestamp_part_1, raw_timestamp_part_2,
                              cvtd_timestamp))
dim(dataset)

# For the test dataset:
dataset_test <- subset(dataset_test[, which(isNA == 0)],
                        select=-c(X, user_name, new_window, num_window,
                                  raw_timestamp_part_1, raw_timestamp_part_2,
                                  cvtd_timestamp))
end.rcode-->

<p>We have reduced the training set to <b>19622 training examples</b> composed of the recording of <b>53 usable measures</b>.</p>


<h1>Step 3: Creation of the training set and validation set</h1>

<p>In order to measure the performances of our future predictor we split the dataset into a training set and a validation set. The first one will be used to train the predictor, while the second one will be used to assess the performances.<br/>
The split used is: <b>80% </b> for the training set and <b>20%</b> for the validation set.</p>

<!--begin.rcode
# Spliting the dataset:
inTrain <- createDataPartition(dataset$classe, p=0.8, list=FALSE)
train_set <- dataset[inTrain,]
valid_set <- dataset[-inTrain,]
end.rcode-->

<h1>Step 4: Training some predictors</h1>

<h2> Test 1: Random Forest:</h1>
<p>We first try an powerful yet easy to parametrize predictor to have an idea of what performance could be obtained. <br/>
We train a <b>Random Forest Classifier</b>.</p>

<!--begin.rcode
# Test 1: Random forest classifier using cross validation control:
ctrl <- trainControl(allowParallel=TRUE, method="cv", number=4)
model <- train(classe ~ ., data=train_set, model="rf", trControl=ctrl)
predictor <- predict(model, newdata=valid_set)
end.rcode-->

<p>Let's now assess the generalization performance of this classifier by computing the error made on the validation set.</p>

<!--begin.rcode
# Error on valid_set:
sum(predictor == valid_set$classe) / length(predictor)
confusionMatrix(valid_set$classe, predictor)$table
end.rcode-->

<p>We have a validation error of <b>99.29%</b>.<br/>
We can consider this classifier as good.</p>

<p>Let's see what are the prediction for the test set.</p>

<!--begin.rcode
# Prediction on the test set:
predict(model, newdata=dataset_test)
end.rcode-->

<p>We have the following prediction: <br/>
<b>  B A B A A E D B A A B C B A E E A B B B </b></p>

<h2> Test 2: SVM on lower dimension training examples</h1>

<p>The random forest appears to be a good predictor but it is quite slow to train on the 53 dimensions vectors of our dataset. <br/>
Let's assess the performance of an SVM trained only on the ten most important parameters. <br/>
We train a <b>SVM Classifier</b>.</p>

<!--begin.rcode
# Most important variable for the ramdom forest predictor:
varImp(model)

# Let us reduce the training dataset to those ten variables
dataset_small <- subset(dataset,
                         select=c(roll_belt, pitch_forearm, yaw_belt,
                                  magnet_dumbbell_y, pitch_belt,
                                  magnet_dumbbell_z, roll_forearm,
                                  accel_dumbbell_y, roll_dumbbell,
                                  magnet_dumbbell_x,classe))
dim(dataset_small)
end.rcode-->

<p>We now have <b>19622 training examples</b> composed of the recording of <b>10 measures</b>. (11th parammeter is the class)</p>

<p>Let's train the SVM on this "small" dataset.</p>

<!--begin.rcode
# Training an SVM on the small dataset
svm <- train(classe ~ ., data=dataset_small[inTrain,], model="svm", trControl=ctrl)
predictor_svm <- predict(svm, newdata=valid_set)
end.rcode-->

<p>Let's now assess the generalization performance of this classifier by computing the error made on the validation set.</p>

<!--begin.rcode
# Error on valid_set:
sum(predictor_svm == valid_set$classe) / length(predictor_svm)
confusionMatrix(valid_set$classe, predictor_svm)$table
end.rcode-->

<p>We have a validation error of <b>98.45%</b>.<br/>
The performance are slightly lower but the training is way faster. <br/>
Depending on the application the trade-off could be interesting</p>

<p>Let's see what are the prediction for the test set.</p>

<!--begin.rcode
# Prediction on the test set:
predict(model, newdata=dataset_test)
end.rcode-->

<p>We have the following prediction: <br/>
<b> B A B A A E D B A A B C B A E E A B B B </b> <br/>
We have the same prediction with the random forest trained on the full dataset and with the SVM trained with the reduced dataset. Again the trade-off "speed of training versus "classification performances" could be interesting depending on the application.</p>

</body>
</html>