WEBVTT

1
00:00:01.000 --> 00:00:01.008
- In machine learning,

2
00:00:01.008 --> 00:00:04.002
one of the best ways to learn more about your data

3
00:00:04.002 --> 00:00:07.005
is by classifying it with what you already know.

4
00:00:07.005 --> 00:00:08.007
Think of it this way.

5
00:00:08.007 --> 00:00:09.006
When I was younger,

6
00:00:09.006 --> 00:00:12.003
I worked for an animal shelter in Chicago.

7
00:00:12.003 --> 00:00:13.008
One of the most difficult jobs

8
00:00:13.008 --> 00:00:17.004
was classifying the breed of incoming dogs.

9
00:00:17.004 --> 00:00:19.007
There are hundreds of different dog breeds,

10
00:00:19.007 --> 00:00:22.003
and most dogs are mixed.

11
00:00:22.003 --> 00:00:23.008
Each time we got a new dog,

12
00:00:23.008 --> 00:00:27.001
we would hold it up to the dogs whose breed we already knew.

13
00:00:27.001 --> 00:00:29.000
Then we'd look at some of the features.

14
00:00:29.000 --> 00:00:30.005
Maybe it was the shape of their face

15
00:00:30.005 --> 00:00:32.001
or the color of their hair.

16
00:00:32.001 --> 00:00:36.002
In a sense, the shelter was classifying the unknown dog

17
00:00:36.002 --> 00:00:39.002
by looking at its nearest neighbor.

18
00:00:39.002 --> 00:00:40.005
Of course, it's not easy to tell

19
00:00:40.005 --> 00:00:44.004
whether a dog was a Boston Terrier or a French bulldog.

20
00:00:44.004 --> 00:00:45.007
The closer the match,

21
00:00:45.007 --> 00:00:48.009
the more likely it was to be classified.

22
00:00:48.009 --> 00:00:49.009
Another way to look at it

23
00:00:49.009 --> 00:00:51.009
is we were trying to minimize the distance

24
00:00:51.009 --> 00:00:55.005
between the unknown dog and the known breeds.

25
00:00:55.005 --> 00:00:57.006
If the features were closely matched,

26
00:00:57.006 --> 00:00:58.009
then there was a short distance

27
00:00:58.009 --> 00:01:03.001
between the unknown dog and its nearest neighbor.

28
00:01:03.001 --> 00:01:05.005
A very common supervised machine learning algorithm

29
00:01:05.005 --> 00:01:10.005
for multi-class classification is called K nearest neighbor.

30
00:01:10.005 --> 00:01:12.005
The algorithm plots new data

31
00:01:12.005 --> 00:01:15.004
and compares it to the data that you already have.

32
00:01:15.004 --> 00:01:16.009
Multi-class classification

33
00:01:16.009 --> 00:01:19.004
is different from binary classification,

34
00:01:19.004 --> 00:01:22.002
because there are more than two dog breeds.

35
00:01:22.002 --> 00:01:26.003
Minimizing the distance is a key part of K nearest neighbor.

36
00:01:26.003 --> 00:01:28.008
The closer you are to your nearest neighbors,

37
00:01:28.008 --> 00:01:31.006
the more likely you are to be accurate.

38
00:01:31.006 --> 00:01:32.009
The most common way to do that

39
00:01:32.009 --> 00:01:35.008
is through something called Euclidean distance.

40
00:01:35.008 --> 00:01:37.007
This is a mathematical formula

41
00:01:37.007 --> 00:01:41.003
that can help see the distance between data points.

42
00:01:41.003 --> 00:01:43.004
Now imagine you had millions of dogs

43
00:01:43.004 --> 00:01:46.005
and you wanted to classify them based on their breed.

44
00:01:46.005 --> 00:01:50.004
To start out, you might want to create two key features.

45
00:01:50.004 --> 00:01:52.001
These will help you classify the dogs

46
00:01:52.001 --> 00:01:53.008
that share the same breed.

47
00:01:53.008 --> 00:01:57.007
These are often called classification predictors.

48
00:01:57.007 --> 00:02:00.006
So let's use their weight and the length of their hair.

49
00:02:00.006 --> 00:02:02.005
Then we'll take these two features

50
00:02:02.005 --> 00:02:05.008
and put them on an x y axis diagram.

51
00:02:05.008 --> 00:02:07.000
This is the same diagram

52
00:02:07.000 --> 00:02:09.009
that you used in geometry in school.

53
00:02:09.009 --> 00:02:13.001
Let's put the length of their hair along the y axis

54
00:02:13.001 --> 00:02:15.008
and their weight along the x axis.

55
00:02:15.008 --> 00:02:19.002
Now take 1,000 labeled dogs for the training set.

56
00:02:19.002 --> 00:02:20.006
This will be like the shelter dogs

57
00:02:20.006 --> 00:02:22.004
where we already knew the breed.

58
00:02:22.004 --> 00:02:23.007
We'll put them on the graph

59
00:02:23.007 --> 00:02:26.000
based on their weight and hair length.

60
00:02:26.000 --> 00:02:29.005
Now let's take our unknown dog and put it on the chart.

61
00:02:29.005 --> 00:02:32.001
You can see that it's not matched with another dog,

62
00:02:32.001 --> 00:02:34.006
but it has a bunch of neighbors.

63
00:02:34.006 --> 00:02:36.009
Let's say we use a K of five.

64
00:02:36.009 --> 00:02:38.007
That means that we'd want to put a circle

65
00:02:38.007 --> 00:02:43.005
around our unclassified dog and its five nearest neighbors.

66
00:02:43.005 --> 00:02:46.007
We can see if the distance of the other dogs is shorter,

67
00:02:46.007 --> 00:02:49.008
we'll get a much more accurate classification.

68
00:02:49.008 --> 00:02:52.007
Now let's look at the five nearest neighbors.

69
00:02:52.007 --> 00:02:54.007
You'll see that three of them are shepherds

70
00:02:54.007 --> 00:02:56.008
and two of them are huskies.

71
00:02:56.008 --> 00:02:58.005
You can be somewhat confident

72
00:02:58.005 --> 00:03:01.007
to classify your unknown dog as a shepherd.

73
00:03:01.007 --> 00:03:04.008
There's also a reasonable chance that it's a husky.

74
00:03:04.008 --> 00:03:05.007
K nearest neighbor

75
00:03:05.007 --> 00:03:08.008
is a very common and powerful machine learning algorithm.

76
00:03:08.008 --> 00:03:11.009
That's because you can do more than just classify dogs.

77
00:03:11.009 --> 00:03:14.003
In fact, it's commonly used in finance

78
00:03:14.003 --> 00:03:15.008
to look for the best stocks,

79
00:03:15.008 --> 00:03:18.003
and even predict future performance.