1
00:00:00,560 --> 00:00:05,840
In this video, we will learn how to create a possible Smardon and Biton.

2
00:00:08,820 --> 00:00:13,660
So let's first import the ACCE boost as X DBE.

3
00:00:16,360 --> 00:00:24,520
So if this Ximo says not installed into your system, you first have to install it to install it.

4
00:00:24,850 --> 00:00:28,420
You have to write up, install extreme boost.

5
00:00:32,200 --> 00:00:33,760
So go to your command prompt.

6
00:00:37,150 --> 00:00:41,340
Use a keyboard shortcut over no art and write CMB and head.

7
00:00:41,440 --> 00:00:41,830
Okay.

8
00:00:44,120 --> 00:00:45,740
And here you can write that.

9
00:00:51,300 --> 00:00:51,650
Doug.

10
00:00:52,830 --> 00:00:54,560
Thanks to boost.

11
00:01:10,890 --> 00:01:13,080
Now, we have installed store loaded cibils.

12
00:01:14,130 --> 00:01:16,560
And if you are using Anaconda.

13
00:01:18,730 --> 00:01:27,000
You can also use an opened up prompt and sort of tip in so you can write KONDA this X reboost.

14
00:01:27,850 --> 00:01:29,110
That will also work.

15
00:01:33,220 --> 00:01:35,720
Let's again import exhibitionist.

16
00:01:39,590 --> 00:01:46,060
So we have important egg symbols and to look at the documentation of symbols.

17
00:01:46,230 --> 00:01:47,550
You can click on this link.

18
00:01:52,760 --> 00:01:56,960
Here you can get all the information you won for excuse.

19
00:01:57,890 --> 00:01:59,720
Let's look at the parameters.

20
00:02:02,150 --> 00:02:05,030
So there are multiple parameters on exit boost.

21
00:02:05,690 --> 00:02:12,980
We can divide this parameters in two, three categories vs. the general parameter, such as learning

22
00:02:12,980 --> 00:02:16,260
rate, learning where it is known as ETR, etc..

23
00:02:17,150 --> 00:02:25,790
And then there are three related parameters, such as maximum BEB Gamma Gamma is similar to your minimum

24
00:02:25,790 --> 00:02:26,690
sample size.

25
00:02:29,160 --> 00:02:32,640
Then subsample Dror.

26
00:02:37,440 --> 00:02:40,460
And then there are learning task parameters.

27
00:02:41,260 --> 00:02:41,490
Oh.

28
00:02:42,030 --> 00:02:50,850
If you remember, the only difference between and will send extra boost is in the regularisation part.

29
00:02:52,440 --> 00:03:00,690
And if you are aware of anyone, then a two regularisation techniques which are laso and reach in case

30
00:03:00,690 --> 00:03:01,770
of linear regression.

31
00:03:03,010 --> 00:03:09,820
There we have shrinkage parameters for the number of variables we use in our model.

32
00:03:13,300 --> 00:03:16,480
So those parameters are Alpha and Lemina.

33
00:03:18,550 --> 00:03:20,700
So there are many people that I went to to see it.

34
00:03:22,290 --> 00:03:27,020
Go through documentation to look at all the parameters that you want to apply.

35
00:03:27,070 --> 00:03:32,190
Will only covering a subset of those parameters which are important.

36
00:03:35,810 --> 00:03:38,990
Let's first create exhibits classifier.

37
00:03:39,110 --> 00:03:41,720
We are using maximum depth of five.

38
00:03:42,320 --> 00:03:45,350
And number of trees as 10000.

39
00:03:45,710 --> 00:03:47,540
And learning rate of point three.

40
00:03:50,360 --> 00:03:51,320
Let's run this.

41
00:03:53,470 --> 00:03:55,060
And let's for our data.

42
00:03:59,420 --> 00:04:01,100
So our model is ready.

43
00:04:01,550 --> 00:04:03,250
Let's get decorously score.

44
00:04:08,630 --> 00:04:10,100
The accuracy is scored.

45
00:04:10,130 --> 00:04:13,030
We are getting a success expunged six six percent.

46
00:04:13,820 --> 00:04:18,530
So, as you can see, we are getting the highest score using Ximo Smardon.

47
00:04:21,700 --> 00:04:30,310
And usually exhibits gifts, the best result out of random forest GBM or any other classified matter.

48
00:04:34,640 --> 00:04:40,130
The Zimbalist, we also get another matter that this plot underscore the importance.

49
00:04:42,430 --> 00:04:48,880
This my card will give us the graph of the relative importance of variables that we are using for our

50
00:04:48,890 --> 00:04:50,920
modern food on this.

51
00:04:54,330 --> 00:04:59,580
You can see we have our variables listed on the left hand side.

52
00:05:00,900 --> 00:05:05,430
And we have the feature importance on this axis.

53
00:05:06,690 --> 00:05:11,790
So as you can see, the most important feature from our variables is time taker.

54
00:05:12,120 --> 00:05:17,340
The second most important feature is Twitter hashtag, then Greller views and so on.

55
00:05:17,520 --> 00:05:21,570
And the least important feature is Johna Underscore Community.

56
00:05:25,090 --> 00:05:27,510
This is another advantage of agcy boost first.

57
00:05:27,620 --> 00:05:29,850
The prediction accuracy is very high.

58
00:05:30,620 --> 00:05:34,640
And you also get the feature importance graph using X equals.

59
00:05:37,710 --> 00:05:46,110
Now, as we discussed earlier, we can use grid search to provide multiple values of our hyper parameters.

60
00:05:47,370 --> 00:05:52,620
So let's use grid search to optimize our exhibits classifier.

61
00:05:53,550 --> 00:05:59,190
So first, I will create just a classifier with Tyber parameters that I don't want to change.

62
00:05:59,940 --> 00:06:07,640
So I will write exhibit classifier with parameters and as symmetries, let's change it to 250 instead

63
00:06:07,650 --> 00:06:16,200
of 500 so that we can quickly drain this model with learning rate zero point one and our numbers set

64
00:06:16,200 --> 00:06:16,830
off 42.

65
00:06:17,490 --> 00:06:20,040
Now we want to change all of this parameters.

66
00:06:20,490 --> 00:06:22,000
First one is the maximum depth.

67
00:06:22,530 --> 00:06:24,060
This the depth of our tree.

68
00:06:24,540 --> 00:06:28,950
And we want to give values from three to 10 with interval of two.

69
00:06:29,190 --> 00:06:32,790
So we want to give three, five, seven and nine.

70
00:06:34,860 --> 00:06:42,810
We can do this either by providing a list or we can also broyd this in the form of fringe method.

71
00:06:43,080 --> 00:06:48,000
So arrange range starting point and point and the step.

72
00:06:49,320 --> 00:06:55,790
So if you are not comfortable with trends, you can also write three Colma, five comma, seven common

73
00:06:55,820 --> 00:06:56,080
name.

74
00:06:57,330 --> 00:06:58,680
Then we have the gala value.

75
00:06:59,040 --> 00:07:03,840
This is almost same as the minimum sample split.

76
00:07:04,230 --> 00:07:06,540
You can read documentation for more detail.

77
00:07:06,780 --> 00:07:09,790
But this is almost same as minimum simple estate.

78
00:07:10,710 --> 00:07:14,400
We are creating value of zero point one zero point one zero one three.

79
00:07:15,240 --> 00:07:18,150
Now we have another parameter, which is subsample.

80
00:07:19,260 --> 00:07:26,480
So if you remember in begging for each tree, we use a subsample of our data.

81
00:07:28,920 --> 00:07:35,850
So we are seeing that old values of 80 percent and 90 percent for subsample value.

82
00:07:36,750 --> 00:07:40,800
That means for each tree use 80 percent of their data.

83
00:07:41,970 --> 00:07:43,690
This 80 percent is random data.

84
00:07:43,740 --> 00:07:46,800
So each time we will select a portion of our data.

85
00:07:48,240 --> 00:07:51,500
Now the next parameter is Colomb subsample by tree.

86
00:07:52,440 --> 00:07:58,380
So this time and sort of data use 80 percent of feature to create ERP.

87
00:07:58,800 --> 00:08:01,220
So randomly we selected two percent of feature.

88
00:08:01,740 --> 00:08:03,570
And we will create a random tree.

89
00:08:04,890 --> 00:08:06,270
Then we have Greg ULFA.

90
00:08:07,080 --> 00:08:10,080
This is the regularization parameter.

91
00:08:10,950 --> 00:08:15,360
And we are providing three values, zero point zero one zero point one and one.

92
00:08:17,810 --> 00:08:22,550
So we are training our model on all these values of fiber parameter.

93
00:08:23,150 --> 00:08:28,100
And after training or model, we will select the best model out of these models.

94
00:08:30,160 --> 00:08:33,530
So let's run this again.

95
00:08:34,000 --> 00:08:40,930
You can notice that we have created all this hyper parameters in the form of dictionary with our hyper

96
00:08:40,930 --> 00:08:43,510
parameter to name as Ghys.

97
00:08:45,070 --> 00:08:48,310
And the range of hyper, I mean, as values.

98
00:08:50,850 --> 00:08:56,790
And again, in grid search, we first have to mention our classifier, which is exhibit underscores

99
00:08:56,790 --> 00:08:57,290
VLF.

100
00:08:57,570 --> 00:08:58,560
Then we have to.

101
00:08:59,770 --> 00:09:00,970
Mentioned the parameters.

102
00:09:01,000 --> 00:09:03,080
We want to pass in the form of dictionary.

103
00:09:03,810 --> 00:09:09,910
We are using padam underscore test one and then cross-validation.

104
00:09:10,140 --> 00:09:13,060
We want cross-validation will to be five.

105
00:09:13,180 --> 00:09:16,540
And we want the scoring to be of accuracy.

106
00:09:18,010 --> 00:09:19,030
Let's run this.

107
00:09:19,420 --> 00:09:23,820
And now we are fitting our Ekstrand in train down to this model.

108
00:09:26,440 --> 00:09:27,580
It will take some time.

109
00:09:28,150 --> 00:09:33,340
And remember that we are using all our models as a classifier models.

110
00:09:34,380 --> 00:09:36,960
All this are available for regression also.

111
00:09:37,260 --> 00:09:41,940
So if you have a continuous very even, you can use that regression.

112
00:09:41,940 --> 00:09:43,680
What is one of these techniques?

113
00:09:49,140 --> 00:09:51,090
Sylvia fitted out grid search.

114
00:09:51,510 --> 00:09:52,320
So let's.

115
00:09:53,720 --> 00:09:54,480
Look at the pet.

116
00:09:54,540 --> 00:09:54,900
I mean.

117
00:09:55,010 --> 00:09:56,120
So for best, Märta.

118
00:09:57,880 --> 00:10:01,420
So we are getting our best mardon when we are using.

119
00:10:02,960 --> 00:10:09,200
The column simple as Boin date, the garment value of point three, maximum depth of seven.

120
00:10:10,850 --> 00:10:16,070
Regularisation as far as zero point one and subsample of zero point eight.

121
00:10:19,920 --> 00:10:25,220
We'll save this more than that is the best modern and do another witty one.

122
00:10:25,260 --> 00:10:27,680
That is Siwiec Zeehan Discourse VLF.

123
00:10:28,710 --> 00:10:30,720
Then we can find out that curacy.

124
00:10:32,070 --> 00:10:35,190
And there scums, so doobies for physics, but say.

125
00:10:36,610 --> 00:10:43,740
Since our data is less and we have used less number of Greece, we are getting our degrees in accuracy

126
00:10:43,790 --> 00:10:44,120
score.

127
00:10:44,230 --> 00:10:52,120
But if you have a large amount of data and if you select large number of Greece, you will definitely

128
00:10:52,120 --> 00:10:59,350
get higher accuracy as compared to a single random XY Wall Street, because here we are picking a best

129
00:10:59,350 --> 00:11:02,020
model out of all this morten's.

130
00:11:04,780 --> 00:11:06,410
That's all for Sewall's.

131
00:11:06,910 --> 00:11:07,300
Thank you.
