﻿1
00:00:00,366 --> 00:00:02,533
Hello and welcome to this art tutorial.

2
00:00:02,533 --> 00:00:05,266
So in the previous tutorial
we learned a very efficient way

3
00:00:05,266 --> 00:00:07,400
to evaluate our model performance.

4
00:00:07,400 --> 00:00:10,400
So that was about evaluating
our model performance.

5
00:00:10,600 --> 00:00:13,300
And in today's tutorial
we are going to learn a technique

6
00:00:13,300 --> 00:00:16,566
that is going to be about
improving model's performance.

7
00:00:16,966 --> 00:00:21,300
So we already built very powerful models
but we can still improve them.

8
00:00:21,300 --> 00:00:22,666
And how can we do that?

9
00:00:22,666 --> 00:00:26,133
Well it's by finding the optimal values
of the hyperparameters.

10
00:00:26,333 --> 00:00:30,300
Because indeed any machine learning model
is composed of two types of parameters.

11
00:00:30,566 --> 00:00:33,566
The first type of parameters
are the parameters that are learned

12
00:00:33,633 --> 00:00:37,233
through the machine learning algorithm,
and the second type of parameters

13
00:00:37,233 --> 00:00:39,000
are the parameters that we choose.

14
00:00:39,000 --> 00:00:43,066
Like for example, the kernel in the SVM
model, or the penalty parameter,

15
00:00:43,200 --> 00:00:47,233
or even some regularization parameter
in ridge regression or lasso,

16
00:00:47,500 --> 00:00:49,533
and basically in any machine
learning model,

17
00:00:49,533 --> 00:00:52,266
we have a lot of parameters
that are not learned.

18
00:00:52,266 --> 00:00:56,433
And so far, what we've done in this
course was just to choose one single value

19
00:00:56,433 --> 00:00:57,366
for these parameters.

20
00:00:57,366 --> 00:01:00,366
We never actually experienced
with these hyperparameters.

21
00:01:00,366 --> 00:01:02,700
We never tried several values of them.

22
00:01:02,700 --> 00:01:06,333
And so that gives us still a lot of room
for improvement, because maybe there were

23
00:01:06,333 --> 00:01:09,333
better choices
for the values of these hyperparameters,

24
00:01:09,366 --> 00:01:12,366
you know, a better choice than the value
we chose.

25
00:01:12,633 --> 00:01:16,466
So this technique called grid Search,
will answer to one of your questions

26
00:01:16,466 --> 00:01:19,466
that we're asked many times
in this course, which is

27
00:01:19,566 --> 00:01:23,266
how do I know which parameter to select
when I make a machine learning model?

28
00:01:23,266 --> 00:01:26,166
What is the optimal value
of these hyperparameters?

29
00:01:26,166 --> 00:01:28,200
And grid
search will give an answer to that

30
00:01:28,200 --> 00:01:31,600
because it will find the optimal values
for these parameters.

31
00:01:31,800 --> 00:01:33,300
So let's see how it will do that.

32
00:01:33,300 --> 00:01:35,566
And let's implement grid search.

33
00:01:35,566 --> 00:01:39,200
So we're going to work on the same problem
as in the previous tutorial.

34
00:01:39,300 --> 00:01:41,200
That is
you know, this classification problem

35
00:01:41,200 --> 00:01:44,800
where we need to classify the users
in the social network and predict

36
00:01:44,966 --> 00:01:47,966
if they're going to click yes or no
on the ad to buy the SUV.

37
00:01:48,066 --> 00:01:51,266
And therefore we are working
with this data set, social network ads.

38
00:01:51,566 --> 00:01:54,966
And besides, that's a good example
because the kernel SVM model

39
00:01:55,200 --> 00:01:56,666
has many parameters.

40
00:01:56,666 --> 00:02:00,000
You know, it has this penalty parameter,
this gamma parameter.

41
00:02:00,300 --> 00:02:03,933
And some of you asked me which values
we should choose for these parameters.

42
00:02:04,233 --> 00:02:06,966
And so grid
search will tell us exactly that.

43
00:02:06,966 --> 00:02:08,533
So let's do it.

44
00:02:08,533 --> 00:02:11,733
We are working in the same folder
as in the previous tutorial.

45
00:02:11,733 --> 00:02:13,566
That is the model selection folder.

46
00:02:13,566 --> 00:02:16,366
And make sure that you have
the social network ad csv file.

47
00:02:16,366 --> 00:02:18,266
And if that's the case you ready to go.

48
00:02:18,266 --> 00:02:21,600
So we are going to apply grid search
anywhere

49
00:02:21,600 --> 00:02:25,900
after the data preprocessing phase,
because we are not actually going to use

50
00:02:26,033 --> 00:02:27,700
this kernel SVM model

51
00:02:27,700 --> 00:02:31,333
and then doing some parameter
tuning on it to find the optimal values.

52
00:02:31,633 --> 00:02:33,900
We are actually going to use a new package

53
00:02:33,900 --> 00:02:37,666
that you will see
is a very practical package to build

54
00:02:37,666 --> 00:02:40,666
the same kernel SVM model
as this one here,

55
00:02:40,666 --> 00:02:43,800
but by applying grid
search on it at the same time.

56
00:02:44,100 --> 00:02:46,500
And this package, you already know it.

57
00:02:46,500 --> 00:02:48,266
It's the carrot package.

58
00:02:48,266 --> 00:02:51,066
And really,
when we talk about machine learning in R,

59
00:02:51,066 --> 00:02:54,233
the carrot package
is one of the most practical packages

60
00:02:54,766 --> 00:02:58,800
because basically with this package
you can build any machine learning model.

61
00:02:58,866 --> 00:03:01,400
And not only you can build
any machine learning model,

62
00:03:01,400 --> 00:03:04,500
but when you build a machine
learning model with the carrot package,

63
00:03:04,700 --> 00:03:08,466
well, it will give you the model
you want with all the optimal parameters.

64
00:03:08,733 --> 00:03:10,433
So that's pretty powerful.

65
00:03:10,433 --> 00:03:13,633
And therefore that's
one of the most powerful packages in R.

66
00:03:14,100 --> 00:03:16,366
So as I told you, we can build

67
00:03:16,366 --> 00:03:20,433
this new tuned model anywhere
after the data preprocessing phase.

68
00:03:20,433 --> 00:03:24,266
So let's actually make it
right after the k fold cross validation

69
00:03:24,266 --> 00:03:28,433
section to, you know, keep
the other interesting code sections here

70
00:03:28,600 --> 00:03:32,600
so that for example, you can compare the
two models, you know, the one we made here

71
00:03:32,833 --> 00:03:35,833
and the new model we were about to make
with the carrot package.

72
00:03:36,100 --> 00:03:38,733
Feel free
to compare their performance results.

73
00:03:38,733 --> 00:03:43,800
So we're going to build this new kernel
SVM model right here.

74
00:03:44,033 --> 00:03:45,100
So applying

75
00:03:45,100 --> 00:03:48,600
grid search to find the best parameters
I prepared this code section here.

76
00:03:48,933 --> 00:03:53,100
But actually it's not only applying grid
search it's also building a new model.

77
00:03:53,566 --> 00:03:58,066
So first the first thing that we have
to do is to import the carrot package.

78
00:03:58,366 --> 00:04:02,066
We already installed it
in some previous tutorials, but let's

79
00:04:02,066 --> 00:04:06,166
make sure you install it if you didn't
follow those previous tutorials.

80
00:04:06,166 --> 00:04:11,000
So I'm just putting back this command
and inside of course carrot in quotes.

81
00:04:11,433 --> 00:04:13,966
All right.
And then I'll put that in command.

82
00:04:13,966 --> 00:04:15,000
Here we go.

83
00:04:15,000 --> 00:04:18,433
And now let's import the package
with the library command.

84
00:04:18,966 --> 00:04:19,700
Here we go.

85
00:04:19,700 --> 00:04:22,866
Library and carrot inside. Perfect.

86
00:04:23,000 --> 00:04:28,100
So now let's build the same kernel
SVM model but with the carrot package.

87
00:04:28,466 --> 00:04:31,133
And you're going to see that it's
going to be something you.

88
00:04:31,133 --> 00:04:31,466
Okay.

89
00:04:31,466 --> 00:04:35,233
So since we're building this
new classifier the kernel SVM classifier.

90
00:04:35,600 --> 00:04:39,000
Well let's go our model
with the usual name that is classifier.

91
00:04:40,233 --> 00:04:40,633
All right.

92
00:04:40,633 --> 00:04:41,866
So that's the variable.

93
00:04:41,866 --> 00:04:42,800
And then equals.

94
00:04:42,800 --> 00:04:45,566
And then that's
where we use the carrot package.

95
00:04:45,566 --> 00:04:47,433
And what is the function
we're going to use. Now.

96
00:04:47,433 --> 00:04:50,266
It's one of the most used function
of the carrot package.

97
00:04:50,266 --> 00:04:53,333
It is the train function. So parentheses.

98
00:04:53,666 --> 00:04:57,566
And now what's really interesting to do
is to take a browser.

99
00:04:58,066 --> 00:04:58,933
Here it is.

100
00:04:58,933 --> 00:05:01,500
And let's type carrot

101
00:05:01,500 --> 00:05:05,800
enter and let's take the GitHub link.

102
00:05:06,133 --> 00:05:07,800
So let's click on it.

103
00:05:07,800 --> 00:05:12,066
And inside of it
you need to click on this link right here.

104
00:05:12,700 --> 00:05:16,133
And then let's click on six
available models.

105
00:05:16,500 --> 00:05:17,966
Because in this section

106
00:05:17,966 --> 00:05:21,666
you'll get all the models
you can build with the carrot package.

107
00:05:21,933 --> 00:05:23,400
And there are a lot of models there

108
00:05:23,400 --> 00:05:25,800
actually
all the models we built in this course.

109
00:05:25,800 --> 00:05:30,066
So right now some of you might be thinking
then why didn't we build all these model

110
00:05:30,066 --> 00:05:32,900
with the carrot package,
since indeed the carrot package

111
00:05:32,900 --> 00:05:35,900
will give us these models
with the best optimal parameters?

112
00:05:36,200 --> 00:05:40,066
Well, that's because the packages we used
to build all the models in the course

113
00:05:40,266 --> 00:05:44,166
have great options, some of which
you cannot use with the carrot package.

114
00:05:44,466 --> 00:05:48,133
So it's good to know how to use both
the packages we used before in the course

115
00:05:48,400 --> 00:05:51,800
and the carrot package,
but definitely for parameter tuning

116
00:05:52,033 --> 00:05:54,300
you should use the carrot package.

117
00:05:54,300 --> 00:05:54,600
All right.

118
00:05:54,600 --> 00:05:56,100
So this list you see

119
00:05:56,100 --> 00:05:59,666
here is the list of all the models
you can build with the carrot package.

120
00:06:00,000 --> 00:06:03,000
And since right now
we are building the kernel

121
00:06:03,000 --> 00:06:06,466
SVM model,
well you will see that it is available.

122
00:06:06,466 --> 00:06:09,466
I mean it is part of this list and

123
00:06:09,500 --> 00:06:13,466
we can find it right here
at the bottom of this list.

124
00:06:13,466 --> 00:06:17,566
Because indeed we can see that we have
many support vector machines model.

125
00:06:17,900 --> 00:06:21,200
And the one we are interested in
is the support

126
00:06:21,200 --> 00:06:24,333
vector machines with radial basis
function kernel.

127
00:06:24,633 --> 00:06:28,200
Remember the radial basis function
kernel is the Gaussian kernel,

128
00:06:28,400 --> 00:06:32,000
the most commonly used kernel
to build the kernel SVM model.

129
00:06:32,366 --> 00:06:36,066
And so this is the model we are going to
build right now with the carrot package.

130
00:06:36,466 --> 00:06:40,366
And the information that we need to take
right now that is going to be one

131
00:06:40,366 --> 00:06:43,233
input of the trained function
we were about to use.

132
00:06:43,233 --> 00:06:45,866
Is this information here?

133
00:06:45,866 --> 00:06:49,433
That's the input of one of the parameters
of the train function.

134
00:06:49,433 --> 00:06:50,933
The method parameter.

135
00:06:50,933 --> 00:06:53,933
And it is with this parameter
that the train function will know

136
00:06:53,933 --> 00:06:56,933
which model to build
and which model to tune.

137
00:06:56,933 --> 00:07:00,766
So right now
what we have to do is to take this name,

138
00:07:01,300 --> 00:07:05,133
copy it, and we will paste it
as input of the method parameter

139
00:07:05,133 --> 00:07:06,900
inside the train function.

140
00:07:06,900 --> 00:07:08,766
So let's go back to RStudio.

141
00:07:08,766 --> 00:07:11,733
Here we go. And now let's build the model.

142
00:07:11,733 --> 00:07:14,400
So we can actually press

143
00:07:14,400 --> 00:07:17,400
F1 here
to get the informations of the arguments.

144
00:07:17,833 --> 00:07:19,900
So let's see what we have to input.

145
00:07:19,900 --> 00:07:24,300
The first compulsory argument
we need to input is this form argument.

146
00:07:24,666 --> 00:07:26,800
And of course that's the formula.

147
00:07:26,800 --> 00:07:28,066
So let's do it.

148
00:07:28,066 --> 00:07:31,600
Let's input
this first argument form equals.

149
00:07:31,600 --> 00:07:33,266
And then we need to put the formula

150
00:07:33,266 --> 00:07:36,300
exactly as we used to do
when building the previous models.

151
00:07:36,533 --> 00:07:41,066
So that's the dependent variable
which is as a reminder purchased.

152
00:07:41,466 --> 00:07:44,466
You know it's the social network
and business problem.

153
00:07:44,533 --> 00:07:47,100
So here we need to input purchased.

154
00:07:48,500 --> 00:07:49,066
Here we go.

155
00:07:49,066 --> 00:07:53,466
Then tilde
and then all the independent variables.

156
00:07:53,833 --> 00:07:56,700
And remember we don't have to put the
names of all the independent variables.

157
00:07:56,700 --> 00:07:59,166
We can use this shortcut here.

158
00:07:59,166 --> 00:08:00,933
That is this dot.

159
00:08:00,933 --> 00:08:03,933
All right then comma
and then next argument.

160
00:08:04,133 --> 00:08:05,800
So the next argument is data.

161
00:08:05,800 --> 00:08:08,433
And that's of course your training set.

162
00:08:08,433 --> 00:08:11,000
You're building your classifier
on a training set.

163
00:08:11,000 --> 00:08:14,000
And so you need to input here data

164
00:08:14,433 --> 00:08:17,133
equals training set.

165
00:08:17,133 --> 00:08:17,900
Here we go.

166
00:08:17,900 --> 00:08:21,633
So now is the form
the formula and the data the training set.

167
00:08:21,933 --> 00:08:24,933
You have all the information
you need to train your model.

168
00:08:25,466 --> 00:08:28,633
But then of course we need to specify
which model we have to build.

169
00:08:28,933 --> 00:08:32,533
That is we need to specify
that we want to make the kernel SVM model.

170
00:08:32,833 --> 00:08:35,633
And that's
what happens with the third argument,

171
00:08:35,633 --> 00:08:38,400
which is not one of these arguments here.

172
00:08:38,400 --> 00:08:42,033
These are not compulsory
but it's the method argument.

173
00:08:42,333 --> 00:08:46,400
Actually you have the link here, the link
we just browsed on Google.

174
00:08:46,400 --> 00:08:50,300
This link will give you the list
of all the models available and correct.

175
00:08:50,766 --> 00:08:53,766
And so that's for this method parameter
that we need to input.

176
00:08:53,800 --> 00:08:57,433
Once we copy it in this link,
that is SVM radio.

177
00:08:57,833 --> 00:08:59,866
So let's input that right now coma.

178
00:08:59,866 --> 00:09:04,500
And then method equals then quotes.

179
00:09:04,866 --> 00:09:08,266
And inside of these quotes
we need to paste SVM radio.

180
00:09:08,800 --> 00:09:12,066
And that's actually all
with these three arguments

181
00:09:12,333 --> 00:09:17,200
the formula the data and the method
you will build a kernel SVM model.

182
00:09:17,466 --> 00:09:20,466
And then for the parameter tuning
you're going to see what happens.

183
00:09:20,500 --> 00:09:25,566
So first what we're going to do
is execute the data preprocessing phase.

184
00:09:25,566 --> 00:09:29,733
Because we need to import the data set
and applying the data preprocessing phase.

185
00:09:30,066 --> 00:09:33,133
So let's do it right now
we're not going to execute these sections

186
00:09:33,133 --> 00:09:36,000
because this is to build
the kernel SVM model the other way.

187
00:09:36,000 --> 00:09:37,833
So we just need
to take everything from here

188
00:09:39,000 --> 00:09:40,433
up to the top.

189
00:09:40,433 --> 00:09:42,133
Here we go. Let's do it.

190
00:09:42,133 --> 00:09:44,166
All right. Data set. Well import it.

191
00:09:44,166 --> 00:09:47,700
We have the data set the training set
and the test set and the data

192
00:09:47,700 --> 00:09:50,700
preprocessing phase was applied
all correctly.

193
00:09:50,800 --> 00:09:51,133
All right.

194
00:09:51,133 --> 00:09:56,433
So now let's build the kernel SVM model
with the carrot package.

195
00:09:56,766 --> 00:09:59,766
So let's make sure we import the carrot
package.

196
00:10:00,133 --> 00:10:00,966
Here we go.

197
00:10:00,966 --> 00:10:05,800
And now let's select this line
and execute it.

198
00:10:06,366 --> 00:10:06,833
All right.

199
00:10:06,833 --> 00:10:08,333
It's taking a little time.

200
00:10:08,333 --> 00:10:10,933
Well around one second that's fine.

201
00:10:10,933 --> 00:10:12,900
And here we get our classifier.

202
00:10:12,900 --> 00:10:15,533
So that's the kernel SVM model classifier.

203
00:10:15,533 --> 00:10:18,533
It is not yet tuned but you're going to
see what's going to happen.

204
00:10:18,566 --> 00:10:22,366
Because now what we're going to do
is pressing enter here.

205
00:10:22,766 --> 00:10:27,300
And we're simply going to type classifier
and selecting this.

206
00:10:27,300 --> 00:10:28,866
And pressing enter.

207
00:10:28,866 --> 00:10:32,466
And there you get
a lot of very interesting informations

208
00:10:32,733 --> 00:10:36,966
about your kernel SVM classifier
you just built with the carrier package.

209
00:10:37,466 --> 00:10:42,333
Because indeed what you see here
is the conclusion of some parameter tuning

210
00:10:42,333 --> 00:10:46,133
that was done on this kernel
SVM classifier you just built.

211
00:10:46,433 --> 00:10:47,966
And what is this conclusion?

212
00:10:47,966 --> 00:10:52,700
It is that accuracy was used to select
the optimal model using the largest value.

213
00:10:52,700 --> 00:10:54,166
So that means that the accuracy

214
00:10:54,166 --> 00:10:57,366
was the performance metric
to evaluate your model performance.

215
00:10:57,966 --> 00:11:01,500
And just below
you see that the final values used for

216
00:11:01,500 --> 00:11:04,966
the model were Sigma equals 1.12,

217
00:11:05,133 --> 00:11:08,400
1.13, and c equals 0.5.

218
00:11:08,966 --> 00:11:11,900
And that's the results
of your parameter tuning.

219
00:11:11,900 --> 00:11:15,600
That's the optimal values
of the hyperparameters of the kernel

220
00:11:15,600 --> 00:11:19,200
SVM model, the ones that will make
your kernel SVM model

221
00:11:19,466 --> 00:11:22,800
even more performance than the one
you built with the previous method,

222
00:11:23,100 --> 00:11:27,066
because of course, maybe the default value
for the C parameter was 0.5.

223
00:11:27,066 --> 00:11:31,100
This looks like the default value,
but definitely the default value

224
00:11:31,100 --> 00:11:34,600
for the sigma parameter, which is another
hyper parameter of kernel.

225
00:11:34,600 --> 00:11:40,133
SVM was definitely not 1.13
and was this 1.13 value.

226
00:11:40,133 --> 00:11:43,800
For the sigma parameter,
you get the optimal accuracy.

227
00:11:44,433 --> 00:11:46,933
An accuracy
that is higher than the accuracy

228
00:11:46,933 --> 00:11:50,000
you obtained with the kernel
SVM model built previously,

229
00:11:50,500 --> 00:11:52,966
and you can even see
what this accuracy is.

230
00:11:52,966 --> 00:11:54,333
It is shown right here.

231
00:11:54,333 --> 00:11:56,966
Accuracy is 92%.

232
00:11:56,966 --> 00:11:58,766
And besides this is a relevant accuracy.

233
00:11:58,766 --> 00:12:01,766
This is not the accuracy
measured on one traintestsplit.

234
00:12:02,033 --> 00:12:04,600
This was measured
on several traintestsplit

235
00:12:04,600 --> 00:12:07,433
which you can exactly see here
with this information.

236
00:12:07,433 --> 00:12:10,433
Resampling bootstrap was 25 repetitions,

237
00:12:10,666 --> 00:12:13,900
which is exactly the same principle
as k fold cross-validation.

238
00:12:14,166 --> 00:12:17,333
That means that you take different samples
of training sets and test sets,

239
00:12:17,600 --> 00:12:19,366
and you build your model
on the training set

240
00:12:19,366 --> 00:12:23,800
and test its performance on the test set
for each of these samples, and eventually

241
00:12:23,800 --> 00:12:28,000
you take the mean of the accuracies
and you get a 92% accuracy.

242
00:12:28,400 --> 00:12:31,100
So that's definitely a good accuracy.

243
00:12:31,100 --> 00:12:34,066
And besides, that's the best
you can obtain with

244
00:12:34,066 --> 00:12:36,900
these optimal values
of the hyperparameters.

245
00:12:36,900 --> 00:12:40,500
Sigma equals 1.13 and c equals 0.5.

246
00:12:40,933 --> 00:12:44,100
And now I'll give you an even faster way
to get these optimal values

247
00:12:44,100 --> 00:12:45,633
of the hyperparameters.

248
00:12:45,633 --> 00:12:49,333
You just need to
well let's keep this classifier here.

249
00:12:49,333 --> 00:12:50,933
But let's copy that.

250
00:12:50,933 --> 00:12:54,900
You just need to take your classifier
and then add a dollar sign.

251
00:12:55,433 --> 00:12:57,500
And then you can add here best tune.

252
00:12:58,533 --> 00:12:59,566
Here you go.

253
00:12:59,566 --> 00:13:02,166
You select it, you execute it.

254
00:13:02,166 --> 00:13:05,933
And you get directly
the optimal values of the hyperparameters

255
00:13:06,300 --> 00:13:10,000
sigma equals 1.13 and c equals 0.5.

256
00:13:10,500 --> 00:13:14,700
And now congratulations
you got your best kernel SVM model

257
00:13:14,966 --> 00:13:18,300
with the best training coefficients
and the best hyperparameters.

258
00:13:18,600 --> 00:13:20,500
And now the choice is yours.

259
00:13:20,500 --> 00:13:24,466
You can either take this kernel
SVM classifier you built here

260
00:13:24,466 --> 00:13:25,766
with the carrier package,

261
00:13:25,766 --> 00:13:29,633
or you can take the classifier
you built here with the usual way.

262
00:13:29,900 --> 00:13:32,900
But inside of this SVM function
you can input

263
00:13:33,000 --> 00:13:36,133
the optimal values
you found here for the hyperparameters.

264
00:13:36,133 --> 00:13:40,066
That is, sigma equals
1.13 and c equals 0.5.

265
00:13:40,466 --> 00:13:41,900
I will let you that for practice

266
00:13:41,900 --> 00:13:44,900
and judge yourself
which method you want to choose.

267
00:13:45,233 --> 00:13:47,966
And I'll see you in the next
and last section of this course

268
00:13:47,966 --> 00:13:49,433
in which we are going to implement

269
00:13:49,433 --> 00:13:53,100
one of the most powerful models
in machine learning, XGBoost.

270
00:13:53,400 --> 00:13:55,833
So this is going to be
a very exciting section.

271
00:13:55,833 --> 00:13:57,300
I can't wait to see you there.

272
00:13:57,300 --> 00:13:59,133
And until then, enjoy machine learning.
