1
00:00:00,780 --> 00:00:07,050
Before we start discussing declassification techniques, we must take a moment to answer the question,

2
00:00:07,500 --> 00:00:10,800
why can't we use linear regression to do classification?

3
00:00:12,090 --> 00:00:13,200
Consider this example.

4
00:00:15,270 --> 00:00:22,500
I have the data set of some profile of people which include whether that's true or not, that great

5
00:00:22,500 --> 00:00:28,090
card balance and their incomes corresponding to each profile.

6
00:00:28,530 --> 00:00:30,390
I have the great risk profile.

7
00:00:30,780 --> 00:00:34,980
That is whether they defaulted on their loan payment or not.

8
00:00:35,490 --> 00:00:38,490
That is whether they repaid their loan or not.

9
00:00:39,690 --> 00:00:42,800
So if they repay, the default value will be no.

10
00:00:43,050 --> 00:00:44,880
If they do not, they are defaulting.

11
00:00:45,330 --> 00:00:47,410
That is their default value will be, yes.

12
00:00:49,370 --> 00:00:54,360
So using this data, I want to classify how risky a particular profiler's.

13
00:00:57,210 --> 00:01:05,010
Now, if I transform this data so that hiders is represented by one that is wherever default values,

14
00:01:05,010 --> 00:01:08,570
yes, I put one and wherever there is lowest.

15
00:01:08,990 --> 00:01:10,620
I represented by minus one.

16
00:01:11,340 --> 00:01:12,990
I can run a linear regression on it.

17
00:01:16,140 --> 00:01:21,850
So with this coding, when I an integration aggression, it will give me some predicted continuous values

18
00:01:21,850 --> 00:01:25,540
of the response variable, and I can say that.

19
00:01:26,800 --> 00:01:32,200
If the predicted value of the response variable is greater than zero, then it belongs to the class

20
00:01:32,200 --> 00:01:34,220
Mondelēz one, which is hiders.

21
00:01:35,300 --> 00:01:41,390
And if it is less than zero, it belongs to the class Magda's minus one, which is less risk.

22
00:01:42,860 --> 00:01:46,880
So this is a fairly good classifying technique, but it has several problems.

23
00:01:47,720 --> 00:01:53,690
The first problem we face with this classifier, which is using linear regression, is if we have more

24
00:01:53,690 --> 00:01:55,950
than two levels in the response variable.

25
00:01:56,600 --> 00:02:00,440
For example, in the response variable instead of yes and no.

26
00:02:00,560 --> 00:02:04,220
If I had credit risk, low, medium and high.

27
00:02:05,270 --> 00:02:07,970
Such a scenario cannot be handled using this.

28
00:02:07,970 --> 00:02:08,840
Classify it.

29
00:02:10,290 --> 00:02:17,640
So as we discussed in dummy variable Clie creation, Lichter, we cannot simply assign zero one two

30
00:02:17,760 --> 00:02:19,380
values to these three classes.

31
00:02:21,000 --> 00:02:23,610
We have to cleared and minus one dummy variables.

32
00:02:24,630 --> 00:02:30,150
In this case, we will have to create three minus one, that is to do me with Abels for credit risk.

33
00:02:32,640 --> 00:02:34,220
Then we will have to response.

34
00:02:34,230 --> 00:02:37,870
We're able to be predicted, which cannot be handled.

35
00:02:38,520 --> 00:02:44,580
Therefore, if response variable has more than two levels, linear regression cannot be used.

36
00:02:47,280 --> 00:02:54,360
Second problem is that the predicted value of life from this classify, it can not be considered as

37
00:02:54,360 --> 00:02:57,380
the probability of Y belonging to a particular class.

38
00:02:59,320 --> 00:03:06,280
Since its values ranging from minus infinity to plus infinity, as you can see by this graph, it's

39
00:03:06,280 --> 00:03:07,030
a straight line.

40
00:03:07,450 --> 00:03:10,760
If you extend it further, it will go up to plus infinity.

41
00:03:11,320 --> 00:03:14,680
And if you extend it backwards, it can go to minus infinity.

42
00:03:16,180 --> 00:03:19,600
But my probability values can only range between zero and one.

43
00:03:20,530 --> 00:03:27,820
So even though it is classifying in two classes, we are not sure about how confident we can be with

44
00:03:27,820 --> 00:03:28,940
deep predictive response.

45
00:03:31,180 --> 00:03:35,590
The third problem we face is the impact of outlaying values.

46
00:03:36,250 --> 00:03:40,100
I'm not talking about the outliers which we handle in Prepossessing.

47
00:03:40,660 --> 00:03:44,470
I'm talking about genuine but large or very small values.

48
00:03:46,830 --> 00:03:51,780
To fit these far away points, regression model will change this slope.

49
00:03:52,320 --> 00:03:56,220
So imagine if I have another point over here.

50
00:03:57,000 --> 00:04:00,810
This line will try to change its slope to accommodate this point.

51
00:04:01,380 --> 00:04:03,570
And this line will move something like this.

52
00:04:05,220 --> 00:04:09,180
And in this process, we will wrongly classify a lot of points.

53
00:04:11,070 --> 00:04:14,470
Does a Linnean model is very sensitive to outline point.

54
00:04:14,880 --> 00:04:18,810
And when outlaying points out present, it gives out wrong predictions.

55
00:04:20,710 --> 00:04:23,790
In the next we do I will tell you about logistic regression.

56
00:04:24,420 --> 00:04:29,250
I will show you these problems of linear regression graphically and show you how logistic regression

57
00:04:29,460 --> 00:04:30,150
overcomes them.
