1
00:00:00,870 --> 00:00:05,130
In this video, we are going to talk about support vector machines.

2
00:00:07,460 --> 00:00:10,110
We are an extension of support to classify it.

3
00:00:11,350 --> 00:00:18,790
As I told you in last video, support vector machines used Col's to be able to create non-linear boundaries.

4
00:00:19,660 --> 00:00:21,130
And what is a common.

5
00:00:22,380 --> 00:00:27,350
Gunnell is simply some functional relationship between two observations.

6
00:00:28,930 --> 00:00:34,570
And as we change this function, we will get different shapes of decision boundaries.

7
00:00:36,590 --> 00:00:44,390
Basically, we try to take the impact of each observation at a particular point, and this impact is

8
00:00:44,390 --> 00:00:45,560
given by a function.

9
00:00:51,390 --> 00:00:53,400
If you take the function to be this.

10
00:00:55,780 --> 00:00:58,330
That is in a product of two points.

11
00:00:58,900 --> 00:01:03,910
You will end up with a straight line by doing some mathematics.

12
00:01:03,970 --> 00:01:11,200
It can be shown that this linear kernel is effectively support better classifier only bartends.

13
00:01:11,260 --> 00:01:12,940
Mathematics will get complicated.

14
00:01:13,180 --> 00:01:16,210
Plus, you do not need it to solve any business problem.

15
00:01:17,350 --> 00:01:20,690
Therefore, we will not be covering the mathematics bar behind it.

16
00:01:29,590 --> 00:01:36,040
Next, if you want a polynomial type of boundry, the cardinal function looks like this.

17
00:01:37,480 --> 00:01:39,520
It is one plus in a product.

18
00:01:39,750 --> 00:01:46,390
Could the power be value of B, will the domain, the flexibility of this polynomial boundary?

19
00:01:47,620 --> 00:01:53,230
B is also known as degree of this polynomial, for our example?

20
00:01:53,620 --> 00:01:57,220
Well, we were not able to create a good linear separator.

21
00:01:57,970 --> 00:02:02,680
You can see that by using a polynomial boundary of sufficiently high degree.

22
00:02:03,310 --> 00:02:05,840
We can classify the observations much better.

23
00:02:11,360 --> 00:02:17,120
Lastly, this radial cardinal is one of the most frequently used cardinal functions.

24
00:02:18,050 --> 00:02:23,120
It is given by this complicated looking formula and this formula.

25
00:02:23,150 --> 00:02:26,720
The first part is a great deal, distance between two points.

26
00:02:31,170 --> 00:02:36,110
And this radial distance is multiplied with a positive, constant gamma.

27
00:02:37,820 --> 00:02:42,290
Now, the distances large or the value of dormice is large.

28
00:02:42,650 --> 00:02:43,910
This product will be large.

29
00:02:44,910 --> 00:02:49,770
And due to this negative sign, the exponential value will be very small.

30
00:02:52,280 --> 00:02:57,830
So even if the distance off point is not much and got my large.

31
00:02:59,040 --> 00:03:01,950
Nearby point will not have much effect.

32
00:03:02,400 --> 00:03:07,140
Only very close point will impact at any given point.

33
00:03:09,550 --> 00:03:14,980
Therefore, large value of gamma will give us very tight margins.

34
00:03:18,370 --> 00:03:25,860
In short, Gama'a defines how much influence a single training example is going to have at any particular

35
00:03:25,870 --> 00:03:26,290
point.

36
00:03:30,290 --> 00:03:35,420
So larger rally of Gamel mean only closer point have an impact.

37
00:03:36,020 --> 00:03:41,230
And farther away point do not have any impact on that at that particular point.

38
00:03:44,480 --> 00:03:49,960
Does garment will also be a hyper barometer, which we need to tune using cross-validation?

39
00:03:50,870 --> 00:03:58,520
And we will try to find out a value of glamour at which edit it will be minimum this link that I showed

40
00:03:58,520 --> 00:03:59,010
you earlier.

41
00:03:59,090 --> 00:03:59,510
Also.

42
00:04:01,910 --> 00:04:05,270
This we can use to visualize the did a deal compelled also.

43
00:04:05,630 --> 00:04:06,530
So let us take it out.

44
00:04:09,990 --> 00:04:13,890
So by default, this has Radil Cardinals elected.

45
00:04:15,420 --> 00:04:16,440
If you scroll down.

46
00:04:18,760 --> 00:04:21,940
You can see the two hyper barometers that you can tune.

47
00:04:22,120 --> 00:04:28,890
One is sea and one is the cardinal think much of it is the same as the Gama that we discussed into religion.

48
00:04:30,940 --> 00:04:35,040
If you jadi well, you'll see that as if you decrease it.

49
00:04:37,950 --> 00:04:47,000
The cost of misclassification decreases, so our classifier will lose more mis classifications.

50
00:04:47,430 --> 00:04:51,570
So basically it is changing the whole area into green.

51
00:04:51,600 --> 00:04:54,930
That is, it is classifying all the points as green.

52
00:04:56,430 --> 00:05:04,790
But if I increase the cost of misclassification, our model will start to classify more accurately and

53
00:05:05,160 --> 00:05:11,330
it will give us some maundering within which all the red points situated.

54
00:05:11,760 --> 00:05:14,460
And outside which we have all the green point.

55
00:05:17,690 --> 00:05:24,350
This connotes sigma, which is the gamma, as I told you, larger rally of Gamma means that there is

56
00:05:24,470 --> 00:05:26,960
more impact of nearby points.

57
00:05:27,590 --> 00:05:30,980
So let us let us reduce the value of gamma first.

58
00:05:33,360 --> 00:05:38,590
As you can see initially, when I have very low value of gamma.

59
00:05:40,260 --> 00:05:44,050
Everything is green since we have more green points alone.

60
00:05:44,280 --> 00:05:45,300
So everything is green.

61
00:05:45,810 --> 00:05:47,950
When I start to increase the value of Gamma.

62
00:05:49,590 --> 00:05:55,170
You can see the small red areas coming up near red points only.

63
00:05:55,860 --> 00:06:03,210
So the points which are near the red observations, these will be classified as red.

64
00:06:03,750 --> 00:06:10,830
Since we are increasing the importance of nearby point only if I continue to increase the importance

65
00:06:10,830 --> 00:06:15,280
of nearby points, I start to get a bigger boundary.

66
00:06:17,570 --> 00:06:28,130
But after a certain level, the impact of all individual points will be sorted, used, that a boundary

67
00:06:28,130 --> 00:06:29,780
will start behaving erratically.

68
00:06:31,010 --> 00:06:40,760
So it is important that we find our optimum values of both these hyper barometers, C and Gamma, so

69
00:06:40,760 --> 00:06:44,540
that we get the correct boundary to do classification.
