1
00:00:00,733 --> 00:00:01,900
Welcome back to the Ultimate.

2
00:00:01,900 --> 00:00:03,900
Data Science course and I am super.

3
00:00:03,900 --> 00:00:04,933
Excited about today.

4
00:00:04,933 --> 00:00:08,100
I've got an incredible tutorial
prepared for you.

5
00:00:08,100 --> 00:00:11,400
We're going to cover off a very important
topic and that is how to build.

6
00:00:11,566 --> 00:00:13,633
Models. Step by step.

7
00:00:13,633 --> 00:00:15,100
I couldn't hold myself.

8
00:00:15,100 --> 00:00:19,333
I had to add that extra line,
step by step to the name of the tutorial

9
00:00:19,400 --> 00:00:22,500
because that is exactly
what we're going to be looking at.

10
00:00:22,866 --> 00:00:24,466
I'm going to give you a framework

11
00:00:24,466 --> 00:00:28,100
for several different methods,
and it's all going to be step by step.

12
00:00:28,266 --> 00:00:31,066
So let's jump straight into it.

13
00:00:31,066 --> 00:00:33,266
Do you remember the good old days
when we had.

14
00:00:33,266 --> 00:00:36,266
One dependent variable
and one independent variable?

15
00:00:36,300 --> 00:00:37,500
Everything was easy

16
00:00:37,500 --> 00:00:41,566
and we just had a simple linear regression
that we had to build.

17
00:00:41,566 --> 00:00:43,633
And everything worked great.

18
00:00:43,633 --> 00:00:46,933
But now in our data we have all. These.

19
00:00:46,933 --> 00:00:48,066
Columns.

20
00:00:48,066 --> 00:00:49,700
Those easy days are gone.

21
00:00:49,700 --> 00:00:51,466
Now all of these columns are.

22
00:00:51,466 --> 00:00:53,133
Potential predictors for.

23
00:00:53,133 --> 00:00:55,066
Our dependent variable.

24
00:00:55,066 --> 00:00:57,233
And there's just so many of them.

25
00:00:57,233 --> 00:00:58,133
And we.

26
00:00:58,133 --> 00:01:01,166
Need to decide
which ones we want to keep and which.

27
00:01:01,166 --> 00:01:03,366
Ones we want to throw out.

28
00:01:03,366 --> 00:01:06,766
And you'll ask
why do we need to throw out columns?

29
00:01:06,766 --> 00:01:08,066
Why do we need to get rid of data?

30
00:01:08,066 --> 00:01:10,100
Why can't we just use everything.

31
00:01:10,100 --> 00:01:11,133
In our model?

32
00:01:11,133 --> 00:01:12,966
Well, I can think of two reasons off.

33
00:01:12,966 --> 00:01:14,000
The top of my head.

34
00:01:14,000 --> 00:01:17,000
Number one is garbage in, garbage out.

35
00:01:17,200 --> 00:01:18,900
If you throw. In a lot. Of stuff into.

36
00:01:18,900 --> 00:01:23,200
Your model,
then your model will not be a good model.

37
00:01:23,200 --> 00:01:26,900
It won't be reliable, it won't be doing
what it's supposed to be doing.

38
00:01:26,900 --> 00:01:30,333
It's going to be a, so to speak, garbage
model.

39
00:01:30,600 --> 00:01:35,033
And number two, at the end of the day,
you're going to have to explain these.

40
00:01:35,033 --> 00:01:36,766
Variables and understand

41
00:01:36,766 --> 00:01:40,500
the not just the math behind them,
but actually what it means that.

42
00:01:41,100 --> 00:01:44,166
Certain variables
predict the behavior of your.

43
00:01:44,166 --> 00:01:45,566
Dependent variable.

44
00:01:45,566 --> 00:01:46,100
And you will.

45
00:01:46,100 --> 00:01:49,566
Have to explain that to your executives,
to your boss, to,

46
00:01:50,100 --> 00:01:51,300
people you're presenting to.

47
00:01:51,300 --> 00:01:52,166
So if you have a.

48
00:01:52,166 --> 00:01:55,800
Thousand variables, it's not going to be
practical to try and explain.

49
00:01:55,800 --> 00:01:56,000
That.

50
00:01:56,000 --> 00:01:58,833
So you want to keep
only the very important.

51
00:01:58,833 --> 00:02:01,833
Ones,
the ones that actually predict something.

52
00:02:01,866 --> 00:02:04,433
So how do we. Construct a model?

53
00:02:04,433 --> 00:02:08,400
This is the process of building the model,
selecting the right variables.

54
00:02:09,133 --> 00:02:10,633
So how do we construct a model.

55
00:02:10,633 --> 00:02:12,600
Well there are five methods.

56
00:02:12,600 --> 00:02:14,966
That we're going to discuss of building.

57
00:02:14,966 --> 00:02:15,733
Models.

58
00:02:15,733 --> 00:02:17,766
Number one is all in.

59
00:02:17,766 --> 00:02:19,166
Number two is backward.

60
00:02:19,166 --> 00:02:20,000
Elimination.

61
00:02:20,000 --> 00:02:22,566
Number three is forward selection.

62
00:02:22,566 --> 00:02:25,500
Number four is bidirectional elimination.

63
00:02:25,500 --> 00:02:28,233
And number. Five is score. Comparison.

64
00:02:28,233 --> 00:02:29,100
We're going to talk about.

65
00:02:29,100 --> 00:02:31,100
Each one of them just now.

66
00:02:31,100 --> 00:02:32,700
Before we do I wanted to say that.

67
00:02:32,700 --> 00:02:35,666
Sometimes you'll hear.
Stepwise regression.

68
00:02:35,666 --> 00:02:40,900
So stepwise regression actually refers
to number two three and four.

69
00:02:41,666 --> 00:02:46,433
because those are like really
the true step by step methods. but.

70
00:02:46,433 --> 00:02:49,466
Sometimes you will hear
people say stepwise stepwise.

71
00:02:49,466 --> 00:02:51,933
Regression in reference.
To just number four.

72
00:02:51,933 --> 00:02:54,433
So they will replace B bidirectional.

73
00:02:54,433 --> 00:02:57,000
Elimination. With stepwise regression.

74
00:02:57,000 --> 00:02:58,266
And that's fine.

75
00:02:58,266 --> 00:03:00,133
That's
that's normal. That's just. Because.

76
00:03:01,233 --> 00:03:02,800
That's the more as you'll see.

77
00:03:02,800 --> 00:03:03,266
From what we.

78
00:03:03,266 --> 00:03:07,066
Discussed, bidirectional elimination
is kind of the more general approach.

79
00:03:07,066 --> 00:03:10,466
And when people say stepwise regression,
they kind of by default

80
00:03:10,900 --> 00:03:14,333
infer imply, imply

81
00:03:14,333 --> 00:03:17,433
that is bi directional elimination
and you have to infer it from there.

82
00:03:18,100 --> 00:03:18,900
okay.

83
00:03:18,900 --> 00:03:21,900
So let's move on to our methods.

84
00:03:21,933 --> 00:03:22,900
Method number one.

85
00:03:22,900 --> 00:03:24,900
All in. It's not a technical term.

86
00:03:24,900 --> 00:03:26,233
I just call it offline.

87
00:03:26,233 --> 00:03:29,100
Basically what it means
is just throw in all your variables.

88
00:03:29,100 --> 00:03:31,133
Something we just discussed
we shouldn't do.

89
00:03:31,133 --> 00:03:33,933
When would you do
that? One is if you have prior. Knowledge.

90
00:03:33,933 --> 00:03:36,600
If you know that these exact variables.

91
00:03:36,600 --> 00:03:37,500
Are the ones are.

92
00:03:37,500 --> 00:03:40,600
Your true predictors,
you don't have to build anything.

93
00:03:40,600 --> 00:03:43,500
You already know that this is the case.

94
00:03:43,500 --> 00:03:43,866
You might.

95
00:03:43,866 --> 00:03:47,066
Know it from domain knowledge,
or you might know it

96
00:03:47,066 --> 00:03:49,866
because you've done this model before,
or somebody just.

97
00:03:49,866 --> 00:03:51,833
Gave you these variables and said, please.

98
00:03:51,833 --> 00:03:54,300
Build a model. Well,
then you don't really have a choice.

99
00:03:54,300 --> 00:03:56,166
You just build the model.

100
00:03:56,166 --> 00:03:56,700
The other one.

101
00:03:56,700 --> 00:03:59,033
Is you have to perhaps, like,

102
00:03:59,033 --> 00:04:02,700
I can't really think of good examples
here, but maybe there's. Some

103
00:04:03,666 --> 00:04:06,133
framework in your company that.

104
00:04:06,133 --> 00:04:08,400
Says that you have to use
these variables, right?

105
00:04:08,400 --> 00:04:09,700
So it's kind of similar to.

106
00:04:09,700 --> 00:04:12,700
Prior knowledge,
but it's not a, it's not. A.

107
00:04:13,200 --> 00:04:14,166
it's not your decision.

108
00:04:14,166 --> 00:04:16,400
It's you might want. To do it differently,

109
00:04:16,400 --> 00:04:20,700
but there is a framework,
you know, like maybe a bank and to.

110
00:04:21,566 --> 00:04:22,000
predict.

111
00:04:22,000 --> 00:04:24,066
Credit, like the likelihood

112
00:04:24,066 --> 00:04:27,066
of somebody defaulting on something
they have to use these specific variables.

113
00:04:27,866 --> 00:04:29,666
once again, I'm not sure in.

114
00:04:29,666 --> 00:04:32,666
Which industries that would be the case,
but perhaps that could be the case.

115
00:04:33,300 --> 00:04:35,966
And number three,
you would use this method if you're.

116
00:04:35,966 --> 00:04:39,500
Preparing for a backward elimination type

117
00:04:39,500 --> 00:04:43,433
of construction of regression,
which is our next type.

118
00:04:43,433 --> 00:04:46,433
So let's move on to backward elimination

119
00:04:46,833 --> 00:04:49,100
okay. So here comes a step by step stuff.

120
00:04:49,100 --> 00:04:53,100
You might you might want to get
your pens out and write these things down.

121
00:04:53,100 --> 00:04:56,366
Because we're going to have a truly step
by step methanol.

122
00:04:56,700 --> 00:04:57,366
All right.

123
00:04:57,366 --> 00:04:59,266
Backward elimination.

124
00:04:59,266 --> 00:05:01,200
First thing step one.

125
00:05:01,200 --> 00:05:03,833
You have to select a significance
level to.

126
00:05:03,833 --> 00:05:04,800
Stay in the model.

127
00:05:04,800 --> 00:05:07,666
So by default we're going to go with 5%.

128
00:05:07,666 --> 00:05:09,600
So 0.05.

129
00:05:09,600 --> 00:05:11,466
And we're going to use it
in the next step.

130
00:05:11,466 --> 00:05:13,166
So at the beginning you.

131
00:05:13,166 --> 00:05:15,633
Decide on this significance level.

132
00:05:15,633 --> 00:05:17,266
Step two you fit the.

133
00:05:17,266 --> 00:05:19,066
Full model with all possible. Predictors.

134
00:05:19,066 --> 00:05:22,200
So you kind of do that all in approach
which we just talked about.

135
00:05:22,466 --> 00:05:24,300
And you put all of your variables.

136
00:05:24,300 --> 00:05:25,333
Into your model.

137
00:05:26,400 --> 00:05:27,400
And now we're going to start.

138
00:05:27,400 --> 00:05:28,600
Getting rid of them.

139
00:05:28,600 --> 00:05:32,333
Step three you considered the predictor
with the highest p value.

140
00:05:32,366 --> 00:05:34,833
So remember those p values
that we talked about.

141
00:05:34,833 --> 00:05:38,233
In for instance in gretl
or in any software you can see them.

142
00:05:38,500 --> 00:05:42,533
So after you've fitted the model you'll
see the one with the highest p value.

143
00:05:42,933 --> 00:05:45,666
So if p is greater

144
00:05:45,666 --> 00:05:49,500
than your significance level
then you go to step four.

145
00:05:49,666 --> 00:05:51,000
And step four.

146
00:05:51,000 --> 00:05:52,800
Is you have to remove that predictor.

147
00:05:52,800 --> 00:05:56,100
So remove basically the variable
that has the highest p value.

148
00:05:56,433 --> 00:05:58,133
And from step. Four you.

149
00:05:58,133 --> 00:05:59,700
Fit the model without. This variable.

150
00:05:59,700 --> 00:06:02,166
So there's a star here because.

151
00:06:02,166 --> 00:06:03,300
I just wanted to remind myself.

152
00:06:03,300 --> 00:06:10,800
To tell you that if you just remove the 
variable, obviously you can't just say.

153
00:06:10,800 --> 00:06:11,300
That okay.

154
00:06:11,300 --> 00:06:12,833
Now now I've got the new model.

155
00:06:12,833 --> 00:06:14,333
You have to actually refit them all.

156
00:06:14,333 --> 00:06:17,766
You have to re recreate the model,
rebuild it with the.

157
00:06:17,766 --> 00:06:20,133
Fewer number of. Variables. So if you had.

158
00:06:20,133 --> 00:06:23,200
Maybe I don't know, 100 variables
and you removed them, one of them

159
00:06:23,200 --> 00:06:25,566
you have 99 now.
Well you have to rebuild it.

160
00:06:25,566 --> 00:06:27,700
So the coefficients
are going to be different.

161
00:06:27,700 --> 00:06:29,700
The constant is going to be different.

162
00:06:29,700 --> 00:06:32,600
And you need to perform that step.

163
00:06:32,600 --> 00:06:35,100
Because once you remove a variable
it affects.

164
00:06:35,100 --> 00:06:38,100
All the other variables
in your whole regression.

165
00:06:38,633 --> 00:06:41,633
And so after step five
you go back to step three.

166
00:06:41,766 --> 00:06:45,800
Once again you look for the variable
with the highest p value in.

167
00:06:45,800 --> 00:06:49,233
Your new model. You take it out,
you remove.

168
00:06:49,266 --> 00:06:51,133
So basically step four you remove it.

169
00:06:51,133 --> 00:06:51,666
You fit the.

170
00:06:51,666 --> 00:06:54,433
Model again with one less variable
and so on.

171
00:06:54,433 --> 00:06:57,933
You keep doing that
until you come to a point where

172
00:06:58,500 --> 00:07:01,500
though
even the variable with the highest p value

173
00:07:01,733 --> 00:07:05,600
that's p value is still less
than your significance level.

174
00:07:05,600 --> 00:07:07,400
So if that condition p is greater.

175
00:07:07,400 --> 00:07:09,833
Than SL is not correct.

176
00:07:09,833 --> 00:07:11,800
Then you don't go to step four anymore.

177
00:07:11,800 --> 00:07:13,433
You go to Finn.

178
00:07:13,433 --> 00:07:16,100
In this case, Finn is the finish.

179
00:07:16,100 --> 00:07:17,666
Your model is ready.

180
00:07:17,666 --> 00:07:21,400
So as soon as all of the variables
that you have left in your.

181
00:07:21,766 --> 00:07:23,333
Model are.

182
00:07:23,333 --> 00:07:24,366
There, p values are.

183
00:07:24,366 --> 00:07:26,200
Less than the significance level.

184
00:07:26,200 --> 00:07:27,666
Your models prepared.

185
00:07:28,866 --> 00:07:31,200
So that's how the backward elimination
method works.

186
00:07:31,200 --> 00:07:32,700
Let's move on to the next one.

187
00:07:32,700 --> 00:07:35,166
Next method is the forward selection.

188
00:07:35,166 --> 00:07:35,733
This is.

189
00:07:35,733 --> 00:07:36,933
It sounds like the opposite

190
00:07:36,933 --> 00:07:40,100
on the picture in the top right
corner does illustrate the opposite.

191
00:07:40,433 --> 00:07:44,566
But it's much more complex
than just simply reversing.

192
00:07:44,866 --> 00:07:45,566
the.

193
00:07:45,566 --> 00:07:48,533
Procedure
you will see that it's it's a much.

194
00:07:48,533 --> 00:07:50,266
Larger. Procedure.

195
00:07:50,266 --> 00:07:52,200
We started with step one.

196
00:07:52,200 --> 00:07:55,233
Select
a significance level to enter the model.

197
00:07:55,766 --> 00:07:58,766
And in this case once again
we're going to select 5%.

198
00:07:59,700 --> 00:08:00,833
Then we go to step two.

199
00:08:00,833 --> 00:08:04,766
We fit all possible
simple regression models.

200
00:08:04,766 --> 00:08:06,900
So we take.

201
00:08:06,900 --> 00:08:08,600
The dependent variable.

202
00:08:08,600 --> 00:08:09,333
And we create a.

203
00:08:09,333 --> 00:08:11,700
Regression model
with every single independent.

204
00:08:11,700 --> 00:08:13,333
Variable that we have.

205
00:08:13,333 --> 00:08:16,433
And then we select out of all those models
we select the. One.

206
00:08:16,633 --> 00:08:18,966
Which has the lowest p value for.

207
00:08:18,966 --> 00:08:20,366
The independent variable.

208
00:08:21,633 --> 00:08:24,266
So you
can already see that that is by in itself.

209
00:08:24,266 --> 00:08:25,200
A lot of. Work.

210
00:08:25,200 --> 00:08:27,466
Then what we do is we move to step three.

211
00:08:27,466 --> 00:08:30,400
We keep this variable that we've just,

212
00:08:30,400 --> 00:08:32,866
chosen, and we fit all.

213
00:08:32,866 --> 00:08:35,466
Other possible models with one extra.

214
00:08:35,466 --> 00:08:39,333
Predictor
added to the one you already have.

215
00:08:39,333 --> 00:08:41,133
So what does that mean?

216
00:08:41,133 --> 00:08:44,533
That means we've selected a, Simple.

217
00:08:45,000 --> 00:08:46,200
Linear regression.

218
00:08:46,200 --> 00:08:47,500
With one variable.

219
00:08:47,500 --> 00:08:50,500
Now we need to construct all possible.

220
00:08:51,500 --> 00:08:54,733
Linear regressions with two variables
where one of those two

221
00:08:54,733 --> 00:08:56,400
variables is the one over you selected.

222
00:08:56,400 --> 00:08:57,033
So basically we.

223
00:08:57,033 --> 00:08:59,066
Add on

224
00:08:59,066 --> 00:09:01,766
all possible
all of the other variables one by. One.

225
00:09:01,766 --> 00:09:04,633
So we choose
okay. Let's add on this variable.

226
00:09:04,633 --> 00:09:07,266
And then let's add on the next one
like but separately.

227
00:09:07,266 --> 00:09:12,066
So we construct all possible
two variable linear regressions.

228
00:09:12,533 --> 00:09:13,766
And just keeping.

229
00:09:13,766 --> 00:09:16,433
Definitely keeping the variable
that. We've already selected.

230
00:09:16,433 --> 00:09:17,666
So what do we do after that.

231
00:09:18,800 --> 00:09:19,466
Out of all.

232
00:09:19,466 --> 00:09:24,333
Of these possible two variable
regressions, we consider the. One.

233
00:09:24,700 --> 00:09:28,633
Where the new variable that we added
had the lowest p value.

234
00:09:29,666 --> 00:09:32,233
So if that p value is.

235
00:09:32,233 --> 00:09:35,733
Less than our significance level
meaning that you know

236
00:09:35,733 --> 00:09:39,000
that variable is a good one,
it's it's, it's a significant variable.

237
00:09:39,000 --> 00:09:40,966
Then we move back to step three.

238
00:09:40,966 --> 00:09:41,700
What does that mean?

239
00:09:41,700 --> 00:09:42,566
It means that.

240
00:09:42,566 --> 00:09:45,300
Now we have a regression
with two variables.

241
00:09:45,300 --> 00:09:47,266
And now we will add a third variable.

242
00:09:47,266 --> 00:09:49,400
We'll try all possible.

243
00:09:49,400 --> 00:09:51,133
Variables that we have left as.

244
00:09:51,133 --> 00:09:52,500
Our third variable.

245
00:09:52,500 --> 00:09:54,733
And then out of all of those models.

246
00:09:54,733 --> 00:09:57,400
With three variables
we will go to step four. And we'll select.

247
00:09:57,400 --> 00:09:58,966
Again the one with the.

248
00:09:58,966 --> 00:10:01,200
Lowest p value for that third variable.

249
00:10:01,200 --> 00:10:02,166
That we added.

250
00:10:02,166 --> 00:10:03,133
And we'll keep doing that.

251
00:10:03,133 --> 00:10:05,500
So basically
we'll keep growing the regression.

252
00:10:05,500 --> 00:10:08,266
Model but not just random.
And it will be. Actually selecting.

253
00:10:08,266 --> 00:10:11,500
Out of the old all of the possible
combinations every single time

254
00:10:12,300 --> 00:10:15,233
and growing it one variable at a time.

255
00:10:16,200 --> 00:10:18,333
And then we will only stop.

256
00:10:18,333 --> 00:10:23,566
When the variable that we've added
it has a p value that is.

257
00:10:23,566 --> 00:10:25,733
Greater than our significance level.

258
00:10:25,733 --> 00:10:27,633
So when. This condition P is less.

259
00:10:27,633 --> 00:10:29,666
Than SL is not true.

260
00:10:29,666 --> 00:10:30,966
Then we don't go to step three.

261
00:10:30,966 --> 00:10:32,500
We finish the regression.

262
00:10:32,500 --> 00:10:33,133
Why will.

263
00:10:33,133 --> 00:10:35,133
Because that variable that we just added.

264
00:10:35,133 --> 00:10:36,666
Is no longer.

265
00:10:36,666 --> 00:10:38,000
Is not. Significant anymore.

266
00:10:38,000 --> 00:10:38,633
And we.

267
00:10:38,633 --> 00:10:39,833
Also know that we selected.

268
00:10:39,833 --> 00:10:41,700
The one with. The lowest p value.

269
00:10:41,700 --> 00:10:44,700
So there's no other variable
that we can add that.

270
00:10:44,900 --> 00:10:47,233
Its p value will be greater.

271
00:10:47,233 --> 00:10:48,766
will be less than SL.

272
00:10:48,766 --> 00:10:51,433
Any, any regression which is from.

273
00:10:51,433 --> 00:10:52,800
Then onwards.

274
00:10:52,800 --> 00:10:56,133
It will the variable, the new variable
will always be insignificant.

275
00:10:56,266 --> 00:10:58,733
And so here we finish the regression.

276
00:10:58,733 --> 00:11:01,133
And the trick here is that you keep.

277
00:11:01,133 --> 00:11:03,566
Not the current model
but the previous one.

278
00:11:03,566 --> 00:11:04,066
And that makes.

279
00:11:04,066 --> 00:11:08,133
Sense because you've just added a variable
which is insignificant.

280
00:11:08,133 --> 00:11:10,900
So what's the point of that variable.
Just move a step back.

281
00:11:10,900 --> 00:11:12,633
So that's how forward selection works.

282
00:11:12,633 --> 00:11:14,100
I know it's a bit confusing,

283
00:11:14,100 --> 00:11:17,100
but just try to wrap your head around
and maybe read through these instructions.

284
00:11:17,100 --> 00:11:18,000
One more time.

285
00:11:18,000 --> 00:11:22,333
It does make a lot of sense when you and
this like kind of picture, what.

286
00:11:22,666 --> 00:11:23,766
Exactly is. Going on.

287
00:11:25,466 --> 00:11:28,766
And next we're moving
on to the bidirectional elimination.

288
00:11:29,100 --> 00:11:30,833
So this one, as you can.

289
00:11:30,833 --> 00:11:33,600
Assume or perhaps guess.

290
00:11:33,600 --> 00:11:34,500
It combines the two.

291
00:11:34,500 --> 00:11:35,333
Step one.

292
00:11:35,333 --> 00:11:38,500
Select
a significance level to stay or to enter.

293
00:11:38,500 --> 00:11:40,800
And a. Significance. Level to. Stay.

294
00:11:40,800 --> 00:11:43,666
So we're going to select in both cases 5%.

295
00:11:43,666 --> 00:11:46,300
But it's up to you what you select.

296
00:11:46,300 --> 00:11:46,866
Step two.

297
00:11:46,866 --> 00:11:49,200
Perform the next. Step of the.

298
00:11:49,200 --> 00:11:51,300
Forward selection.

299
00:11:51,300 --> 00:11:53,433
meaning that the one that.

300
00:11:53,433 --> 00:11:54,433
We just discussed.

301
00:11:54,433 --> 00:11:58,866
So where new variables, when they enter
in order to enter they have to be.

302
00:11:59,166 --> 00:12:01,333
Less than the significance.

303
00:12:01,333 --> 00:12:03,033
Level two enter.

304
00:12:03,033 --> 00:12:07,266
So basically add on a new variable
based on the forward selection method.

305
00:12:07,700 --> 00:12:11,300
Step three perform all of the steps
of the backward elimination process.

306
00:12:11,300 --> 00:12:12,333
So now.

307
00:12:12,333 --> 00:12:15,866
If you have two variables, start getting
rid of them and see if you can get rid of.

308
00:12:15,866 --> 00:12:17,400
Any of them. And then

309
00:12:18,400 --> 00:12:19,500
move back to step two.

310
00:12:19,500 --> 00:12:21,433
So then grow it by another variable.

311
00:12:21,433 --> 00:12:23,233
And every time you grow it by a variable,
say let's

312
00:12:23,233 --> 00:12:26,233
let's say you were at, five variables
and you went to six.

313
00:12:26,466 --> 00:12:29,566
Since you went to six, you have to from
all of the steps of backward elimination.

314
00:12:29,566 --> 00:12:30,966
So you don't just. Eliminate one variable.

315
00:12:30,966 --> 00:12:32,466
If you can eliminate one, two.

316
00:12:32,466 --> 00:12:34,633
Three, however many you can.

317
00:12:34,633 --> 00:12:37,266
And then from there
you can go back to step two.

318
00:12:37,266 --> 00:12:39,133
And so this is a very iterative process.

319
00:12:39,133 --> 00:12:40,200
You keep doing that.

320
00:12:40,200 --> 00:12:42,333
Until at some. Point you cannot.

321
00:12:42,333 --> 00:12:43,666
Add new variables.

322
00:12:43,666 --> 00:12:46,833
No variables can enter
or no old variables can exit.

323
00:12:46,833 --> 00:12:50,733
And as soon as you get there,
then you proceed to the finish because.

324
00:12:50,733 --> 00:12:52,400
Your model is. Ready.

325
00:12:52,400 --> 00:12:54,466
You can't you can't add anything.
You can't take anything out.

326
00:12:54,466 --> 00:12:57,000
That means you've,
you've created the model.

327
00:12:57,000 --> 00:12:58,833
So this. Is one of the more.

328
00:12:58,833 --> 00:13:00,433
Tedious methods.

329
00:13:00,433 --> 00:13:02,066
Of course, you would have to,

330
00:13:02,066 --> 00:13:06,033
get a computer to do this for you,
because otherwise you'd go go insane.

331
00:13:06,400 --> 00:13:08,300
Putting variables in and taking them out.

332
00:13:08,300 --> 00:13:11,266
But that's
how bidirectional elimination works.

333
00:13:11,266 --> 00:13:15,233
and once again, some people call it

334
00:13:15,533 --> 00:13:20,500
stepwise regression
and finally all possible models.

335
00:13:20,766 --> 00:13:21,666
So this is.

336
00:13:21,666 --> 00:13:27,566
The most probably thorough approach,
but also the most, resource.

337
00:13:27,566 --> 00:13:31,400
Consuming approach because you select
a criterion of goodness of fit.

338
00:13:31,400 --> 00:13:34,400
For instance, like chi key
criterion can be the R squared.

339
00:13:34,566 --> 00:13:35,533
lots of different criteria.

340
00:13:35,533 --> 00:13:38,600
And then you construct
all possible regression model.

341
00:13:38,600 --> 00:13:40,533
So if you had. And variables then.

342
00:13:40,533 --> 00:13:42,033
There will be a two.

343
00:13:42,033 --> 00:13:44,966
To the power of n minus one
total combinations.

344
00:13:44,966 --> 00:13:45,700
Of these variables.

345
00:13:45,700 --> 00:13:49,233
And that that's exactly
how many models there can possibly be.

346
00:13:50,100 --> 00:13:51,066
And then.

347
00:13:51,066 --> 00:13:53,266
Step three
you select the. One of these models.

348
00:13:53,266 --> 00:13:55,500
With the best criterion.
That you're looking at.

349
00:13:56,700 --> 00:13:57,133
There you go.

350
00:13:57,133 --> 00:13:58,200
Your model is ready.

351
00:13:58,200 --> 00:14:01,733
So sounds easy,
but let's have a look at an example.

352
00:14:01,733 --> 00:14:02,166
Even if you have.

353
00:14:02,166 --> 00:14:05,433
Ten columns in your data,
you'll have 1023 models.

354
00:14:05,433 --> 00:14:08,000
That's insane.
That's an insane amount of models.

355
00:14:08,000 --> 00:14:09,800
And we're not talking about columns that.

356
00:14:09,800 --> 00:14:10,966
You've already filter. Out.

357
00:14:10,966 --> 00:14:15,000
So columns that, you, you know,
that's like in our example,

358
00:14:15,000 --> 00:14:18,000
you might have 5 or 6 columns.

359
00:14:18,000 --> 00:14:21,000
Now we're talking about
when you get a data set that you.

360
00:14:21,633 --> 00:14:23,433
You just it's pretty much raw.

361
00:14:23,433 --> 00:14:27,100
And it has like maybe 100 columns
like I've worked with

362
00:14:27,100 --> 00:14:29,200
data sets were around that.

363
00:14:29,200 --> 00:14:32,500
Maybe 50 to 1. Hundred,
maybe more columns.

364
00:14:32,700 --> 00:14:35,066
And instead of going through them,
this is what this.

365
00:14:35,066 --> 00:14:36,200
Method is suggesting.

366
00:14:36,200 --> 00:14:38,266
Instead of going through them
and picking out the ones that.

367
00:14:38,266 --> 00:14:40,733
You think should be in your model,
you just throw. Everything in.

368
00:14:40,733 --> 00:14:43,133
Well, it's not a. Good approach because

369
00:14:44,133 --> 00:14:45,600
basically it's.

370
00:14:45,600 --> 00:14:46,866
it's growing exponentially.

371
00:14:46,866 --> 00:14:49,200
The number of models is growing
exponentially.

372
00:14:49,200 --> 00:14:54,633
And, it's, It's very resource
consuming to get.

373
00:14:54,633 --> 00:14:56,533
A result from this approach.

374
00:14:56,533 --> 00:14:57,900
And finally.

375
00:14:57,900 --> 00:15:00,466
So where have we come to?

376
00:15:00,466 --> 00:15:01,933
We've come to our conclusion.

377
00:15:01,933 --> 00:15:06,333
We have five methods of building
model models, all in backward elimination.

378
00:15:06,333 --> 00:15:06,966
For selection.

379
00:15:06,966 --> 00:15:09,866
Bidirectional elimination
and score comparison.

380
00:15:09,866 --> 00:15:14,700
So the one we're going to be looking
at in these tutorials, in order to get.

381
00:15:14,700 --> 00:15:16,000
Our head around.

382
00:15:16,000 --> 00:15:19,600
How to build models step by step
and get some practice, is the backward.

383
00:15:19,600 --> 00:15:21,933
Elimination. Process because it is the.

384
00:15:21,933 --> 00:15:23,633
Fastest one at all of them.

385
00:15:23,633 --> 00:15:26,833
And you will still get to see
exactly how the step by.

386
00:15:26,833 --> 00:15:28,033
Step method works.

387
00:15:28,033 --> 00:15:28,733
And plus, we'll.

388
00:15:28,733 --> 00:15:31,833
Throw in a few extra tricks
along the way to make sure.

389
00:15:31,833 --> 00:15:34,833
Our models are very robust.

390
00:15:35,166 --> 00:15:36,800
Can't wait to get started!

391
00:15:36,800 --> 00:15:37,800
Lots of theory.

392
00:15:37,800 --> 00:15:40,466
Let's get to the practice.
I'll look forward to seeing you next time.