﻿1
00:00:00,966 --> 00:00:05,800
Welcome to section 34 Principal Component
Analysis, or PCA.

2
00:00:06,366 --> 00:00:09,400
Before you get some hands on experience
coming up with PCA,

3
00:00:09,800 --> 00:00:14,433
in the practical exercises, we're going
to take a look at the intuition behind it.

4
00:00:14,933 --> 00:00:17,466
PCA is considered to be one of the most

5
00:00:17,466 --> 00:00:21,400
used unsupervised
algorithms, and can be seen

6
00:00:21,633 --> 00:00:24,833
as the most popular dimensionality
reduction algorithm.

7
00:00:26,700 --> 00:00:27,833
PCA is used

8
00:00:27,833 --> 00:00:32,133
for operations
such as visualization, feature extraction,

9
00:00:32,133 --> 00:00:36,000
noise filtering, and can be seen in
algorithms used for stock market

10
00:00:36,000 --> 00:00:39,833
predictions and gene analysis,
just to name a few.

11
00:00:42,900 --> 00:00:44,200
The goal of PCA

12
00:00:44,200 --> 00:00:48,000
is to identify
and detect correlation between variables.

13
00:00:48,366 --> 00:00:51,166
If there's a strong correlation
and it's found,

14
00:00:51,166 --> 00:00:54,866
then you could reduce the dimensionality,
which really what PCA is intended for.

15
00:00:55,200 --> 00:00:57,400
You find the directions

16
00:00:57,400 --> 00:00:59,733
of maximum variance
in high dimensional data,

17
00:00:59,733 --> 00:01:03,366
and then you project it into a smaller
dimensional subspace while retaining

18
00:01:03,366 --> 00:01:06,600
most of the information, usually again
with PCA.

19
00:01:06,600 --> 00:01:11,500
The goal to reduce the dimensions of a
deep dimensional data set by projecting it

20
00:01:11,700 --> 00:01:15,866
onto a k dimensional subspace
where K is less than D,

21
00:01:16,733 --> 00:01:20,500
and for a overall breakdown
and wrap up of the PCA,

22
00:01:20,633 --> 00:01:24,600
we can see here that the mean functions

23
00:01:24,600 --> 00:01:27,833
of the PCA algorithm
would be to standardize the data,

24
00:01:29,400 --> 00:01:32,400
obtain the eigenvectors and eigenvalues

25
00:01:32,933 --> 00:01:35,933
to then
sort the eigenvalues in descending order.

26
00:01:37,600 --> 00:01:39,800
Construct the projection matrix W

27
00:01:39,800 --> 00:01:42,800
from the selected k eigenvectors

28
00:01:43,633 --> 00:01:46,100
and to transform the original data set,

29
00:01:46,100 --> 00:01:48,733
and you can explore it further
if you follow that link.

30
00:01:48,733 --> 00:01:51,400
But one thing I want to examine here

31
00:01:51,400 --> 00:01:54,500
with PCA and I think the visualization
is really helpful

32
00:01:55,366 --> 00:01:58,266
if we visit the following link, it's
going to take us to this page

33
00:01:58,266 --> 00:02:02,000
where we can actually view it in 2D
and 3D examples.

34
00:02:02,033 --> 00:02:05,333
Now with PCA in a 2D,
you can start to see the relationship

35
00:02:05,333 --> 00:02:08,333
and how PCA is playing out
amongst the variables in the data.

36
00:02:09,033 --> 00:02:11,366
You can also on this site drag them

37
00:02:11,366 --> 00:02:14,366
around, drag the data points around

38
00:02:14,666 --> 00:02:15,766
to see the PCA.

39
00:02:15,766 --> 00:02:18,600
The coordinates,
adjust within the system.

40
00:02:18,600 --> 00:02:23,800
But really what I think is helpful
is the 3D example with the 3D example,

41
00:02:23,866 --> 00:02:28,200
you can actually see the relationship,
the data within this model

42
00:02:28,600 --> 00:02:31,666
and comparing it to the 2D, you know,
within the higher dimensional space,

43
00:02:31,700 --> 00:02:35,566
obviously it can be a much easier
visualization.

44
00:02:35,800 --> 00:02:39,366
And I think this is helpful
really to kind of grasp what PCA is doing.

45
00:02:39,800 --> 00:02:42,800
And if we drag
the data points around again,

46
00:02:43,166 --> 00:02:46,733
just for a test, we can click on the show

47
00:02:46,800 --> 00:02:49,966
PCA to reset that and we'll show it.

48
00:02:51,433 --> 00:02:53,400
And we can see the PCA here.

49
00:02:53,400 --> 00:02:55,533
We can actually move the model around.

50
00:02:55,533 --> 00:02:59,033
Since it's not on that 2D plot,
we can visualize it within the 3D example.

51
00:02:59,466 --> 00:02:59,733
All right.

52
00:02:59,733 --> 00:03:02,733
So to wrap PCA up short and sweet

53
00:03:02,966 --> 00:03:06,100
PCA is not like linear regression,
although it may look like it

54
00:03:06,800 --> 00:03:11,966
because rather than attempting to predict
the values, PCA is attempting

55
00:03:11,966 --> 00:03:15,266
to learn about the relationship
between the x and y values.

56
00:03:15,633 --> 00:03:18,833
It's quantified
by finding a list of principal axes.

57
00:03:19,833 --> 00:03:22,300
And I think one of the best ways
is to look at the visualizations.

58
00:03:22,300 --> 00:03:24,800
You don't compare it to 2D and 3D
dimension,

59
00:03:24,800 --> 00:03:28,033
the analysis and the visualizations
that we looked at before.

60
00:03:28,500 --> 00:03:32,200
In addition, you know, on a side
note, PCA does have a weakness.

61
00:03:32,200 --> 00:03:37,833
It is highly affected by outliers
in the data, but PCA is considered

62
00:03:37,833 --> 00:03:41,266
to be one of the most used,
and it's extremely popular to use.

63
00:03:41,266 --> 00:03:44,266
And I think once you start working through
the practical, it will make more sense.

64
00:03:44,700 --> 00:03:47,933
If you have any questions, please
let us know and enjoy machine learning.
