﻿1
00:00:00,100 --> 00:00:02,233
Ordinary least squares.

2
00:00:02,233 --> 00:00:04,500
So we have our data points.

3
00:00:04,500 --> 00:00:07,600
Now the question is 
that we're answering in this tutorial is

4
00:00:07,933 --> 00:00:11,366
how do we know
which of the sloped lines is the best one?

5
00:00:11,400 --> 00:00:13,866
Is it
this one? Is it this one? Is it this one?

6
00:00:13,866 --> 00:00:16,000
Is it this one or is it this one?

7
00:00:16,000 --> 00:00:18,766
As we can see,
there can be multiple slope lines

8
00:00:18,766 --> 00:00:21,766
that we can draw through our data points.

9
00:00:21,866 --> 00:00:25,000
And how do we know
which one is the best one.

10
00:00:25,000 --> 00:00:27,000
Which is the best linear regression.

11
00:00:27,000 --> 00:00:30,000
And in fact
how do we even define the best one.

12
00:00:30,466 --> 00:00:32,900
So in order to answer those questions,
we need to look at a method

13
00:00:32,900 --> 00:00:35,000
called the ordinary least squares method.

14
00:00:35,000 --> 00:00:39,333
And the way it works in a visual sense
is we need to take our data points

15
00:00:39,533 --> 00:00:42,533
and project them vertically onto,

16
00:00:42,733 --> 00:00:45,366
our linear regression line.

17
00:00:45,366 --> 00:00:48,433
Now we would need to do this
for every single linear regression line

18
00:00:48,433 --> 00:00:49,500
that we're considering.

19
00:00:49,500 --> 00:00:50,866
But for simplicity's sake, in

20
00:00:50,866 --> 00:00:53,866
this tutorial we're just going to do it
with this, line here in the middle.

21
00:00:54,700 --> 00:00:59,666
Now for each pair of points
we have two values y I and y I hat.

22
00:00:59,966 --> 00:01:01,266
So what are these values?

23
00:01:01,266 --> 00:01:04,833
Y is the actual, amount

24
00:01:05,200 --> 00:01:10,500
of, potatoes in our case, in our example,
potatoes yielded from the farm

25
00:01:11,100 --> 00:01:15,700
when that, specific amount of nitrogen
fertilizer was used.

26
00:01:15,700 --> 00:01:18,800
So let's say 15kg of nitrogen
fertilizer were used

27
00:01:19,033 --> 00:01:22,033
and the farm yielded two tons of potatoes.

28
00:01:22,166 --> 00:01:27,000
Y I hat, on the other hand,
is what this linear regression

29
00:01:27,000 --> 00:01:30,066
that we're considering what it predicts
the yield,

30
00:01:30,600 --> 00:01:33,800
to be or two would have been.

31
00:01:33,800 --> 00:01:38,233
So in this case, let's say I
again, 15kg of nitrogen were used.

32
00:01:38,233 --> 00:01:42,366
But the, linear regression
that we're looking at predicts that only

33
00:01:42,600 --> 00:01:47,866
one and a half tonnes of potatoes,
would have been yielded from the farm.

34
00:01:48,266 --> 00:01:52,400
As you can see,
there's a slight difference between the,

35
00:01:53,066 --> 00:01:56,700
actual value, actual yield
and the predicted.

36
00:01:56,733 --> 00:01:57,633
And that's normal.

37
00:01:57,633 --> 00:01:59,233
It's never going to be perfectly.

38
00:01:59,233 --> 00:02:02,133
This line's never going to go perfectly
through every single data point.

39
00:02:02,133 --> 00:02:04,266
That's simply impossible.

40
00:02:04,266 --> 00:02:07,433
but what we want to do
is we want to find the best line,

41
00:02:07,433 --> 00:02:11,133
and it will be related to how small

42
00:02:11,133 --> 00:02:14,133
these differences are, as we can imagine.

43
00:02:14,233 --> 00:02:15,866
So let's have a look here.

44
00:02:15,866 --> 00:02:17,733
There's our y and y hat.

45
00:02:17,733 --> 00:02:20,033
The difference
between them is called the residual.

46
00:02:20,033 --> 00:02:21,566
here's our equation.

47
00:02:21,566 --> 00:02:27,533
And the best equation is such equation
A where b or where

48
00:02:27,633 --> 00:02:31,466
such an equation where the parameters b0
and b1 are such

49
00:02:31,466 --> 00:02:36,700
that the sum
of the squares of the residuals

50
00:02:37,033 --> 00:02:40,900
is minimized, and that's why it's called
the ordinary least squares method.

51
00:02:40,900 --> 00:02:44,866
So we need to take all of these, 
residuals, these differences.

52
00:02:44,866 --> 00:02:48,900
We need to square them for every single 
data point.

53
00:02:49,400 --> 00:02:54,900
and then we need to add up the sum
and whenever we find the smallest value.

54
00:02:54,900 --> 00:02:58,066
So for whichever regression line,
this value is going to be the smallest,

55
00:02:58,066 --> 00:02:59,500
that will be the best regression line.

56
00:02:59,500 --> 00:03:02,433
And that will guarantee that the line
is going nicely through our data points.

57
00:03:02,433 --> 00:03:05,666
And it is the best line or the best linear

58
00:03:05,666 --> 00:03:09,333
regression
to use for modeling our problem.

59
00:03:09,333 --> 00:03:10,000
So there you go.

60
00:03:10,000 --> 00:03:12,800
That's in a nutshell how the ordinary
least squares method works.

61
00:03:12,800 --> 00:03:14,466
And I look forward
to seeing you next time.

62
00:03:14,466 --> 00:03:16,100
And till then enjoy machine learning.
