﻿1
00:00:00,500 --> 00:00:03,066
Hello and welcome back to the course
on Machine Learning.

2
00:00:03,066 --> 00:00:04,100
Today we're going to cover off

3
00:00:04,100 --> 00:00:07,100
some additional comments
about the Naive Bayes classifier.

4
00:00:07,233 --> 00:00:08,400
All right so let's have a look.

5
00:00:08,400 --> 00:00:09,900
Today we're going to go through things.

6
00:00:09,900 --> 00:00:12,933
Number one is the question
why is it called naive.

7
00:00:13,433 --> 00:00:17,200
Number two is the p of x and how we can
potentially drop it sometimes.

8
00:00:17,700 --> 00:00:19,466
I'll show you a quick shortcut.

9
00:00:19,466 --> 00:00:20,900
And number three is what happens

10
00:00:20,900 --> 00:00:23,966
when there's more than two features
involved in your data set.

11
00:00:24,466 --> 00:00:26,400
All right so let's get started.

12
00:00:26,400 --> 00:00:31,266
Question why is this algorithm
called the Naive Bayes algorithm?

13
00:00:31,733 --> 00:00:33,300
Well, the answer is pretty simple.

14
00:00:33,300 --> 00:00:37,166
The answer is because the Bayes
theorem requires

15
00:00:37,166 --> 00:00:41,333
some independence assumptions,
and the Bayes theorem is the foundation

16
00:00:41,333 --> 00:00:46,466
of the Naive Bayes machine learning
algorithm, and therefore the Naive Bayes.

17
00:00:46,700 --> 00:00:47,900
A machine learning algorithm

18
00:00:47,900 --> 00:00:52,300
also relies on these assumptions,
which are oftentimes not correct,

19
00:00:52,800 --> 00:00:56,533
and therefore it's kind of naive to assume
that they're going to be correct.

20
00:00:56,800 --> 00:00:57,900
That's the reason why it's called.

21
00:00:57,900 --> 00:01:00,900
Let's go back to our example
and see what what that means.

22
00:01:01,066 --> 00:01:03,333
So here we've got age and salary right.

23
00:01:03,333 --> 00:01:04,333
And based on

24
00:01:04,333 --> 00:01:08,166
those we're using the Naive Bayes
algorithm to classify our data points

25
00:01:08,166 --> 00:01:11,333
into people who drive to work
or people who walk to work.

26
00:01:11,933 --> 00:01:15,700
Well, the Bayes theorem,
the way we apply it, actually requires

27
00:01:15,700 --> 00:01:19,533
that age and salary
have to be independent, or the variables

28
00:01:19,533 --> 00:01:22,633
that we're working with in this case
as salary, have to be independent.

29
00:01:23,200 --> 00:01:26,566
And that is like a fundamental assumption
of the Bayes theorem.

30
00:01:26,566 --> 00:01:27,933
And then you can only apply it.

31
00:01:27,933 --> 00:01:30,833
And then you can get those probabilities
and so on.

32
00:01:30,833 --> 00:01:33,866
But in our case,
if you just think about it

33
00:01:33,866 --> 00:01:37,700
fundamentally,
it is probably not the case.

34
00:01:37,700 --> 00:01:41,100
I probably if there is some sort
of correlation between age and salary,

35
00:01:41,100 --> 00:01:45,200
because as a person gets older,
their experience grows,

36
00:01:45,200 --> 00:01:48,366
or the number of years
that they've spent in the workforce grows

37
00:01:48,533 --> 00:01:50,166
and therefore their salary grows.

38
00:01:50,166 --> 00:01:52,133
So it's natural for the salary to grow.

39
00:01:52,133 --> 00:01:55,333
If age, it might not be a super strong
correlation, for it's

40
00:01:55,333 --> 00:01:56,533
not it's not for everybody.

41
00:01:56,533 --> 00:01:59,033
But overall
there is some sort of correlation.

42
00:01:59,033 --> 00:02:01,133
So there are not absolutely
independent variables.

43
00:02:01,133 --> 00:02:02,766
And we're all you can even see
that from the chart.

44
00:02:02,766 --> 00:02:04,466
You can just by looking at our chart,

45
00:02:04,466 --> 00:02:07,800
see that there's some sort of correlation
between the two variables.

46
00:02:08,366 --> 00:02:11,133
And therefore given
that they're not independent,

47
00:02:11,133 --> 00:02:14,166
you can't really apply the Bayes theorem
and therefore you can't

48
00:02:14,166 --> 00:02:17,333
apply the base
algorithm to machine learning.

49
00:02:17,333 --> 00:02:21,400
And that's why it's called Naive Bayes
algorithm, because oftentimes it's applied

50
00:02:21,633 --> 00:02:24,500
even though the variables or the features

51
00:02:24,500 --> 00:02:27,766
are not independent
or not completely independent,

52
00:02:28,200 --> 00:02:30,900
and it's still applied
and it's still gives good results.

53
00:02:30,900 --> 00:02:34,100
And that's why it's called naive
because it's a naive assumption.

54
00:02:34,766 --> 00:02:35,566
All right.

55
00:02:35,566 --> 00:02:37,300
Number £0.02 of x.

56
00:02:37,300 --> 00:02:40,300
So let's have a look at what we performed.

57
00:02:40,366 --> 00:02:43,300
So kind of like rewind and and now analyze

58
00:02:43,300 --> 00:02:46,333
what we did in our steps
in the previous tutorial.

59
00:02:46,333 --> 00:02:50,400
So in step two
what we did is we took p of x.

60
00:02:50,400 --> 00:02:55,300
So when we were calculating p of x we
drew a circle around our new data point.

61
00:02:55,300 --> 00:02:58,200
We removed the data point
just so it's not in the way.

62
00:02:58,200 --> 00:02:59,966
And then we shaded the area.

63
00:02:59,966 --> 00:03:01,800
And so what is p of x.

64
00:03:01,800 --> 00:03:06,033
Well p of x
is the likelihood that a randomly selected

65
00:03:06,033 --> 00:03:09,900
point from this data set
will exhibit the features

66
00:03:09,966 --> 00:03:13,266
similar to the data point
that we were about to add.

67
00:03:13,600 --> 00:03:14,933
And as we agreed,

68
00:03:14,933 --> 00:03:18,833
anything in that circle
is deemed to be similar to our data point.

69
00:03:19,166 --> 00:03:23,300
Another way to think about
it is what happens if I throw in a random

70
00:03:23,300 --> 00:03:27,233
variable or a random data
point into this data set right now,

71
00:03:27,666 --> 00:03:30,800
what is the likelihood
that it will fall into the circle?

72
00:03:30,800 --> 00:03:34,400
What is the likelihood
that it will exhibit features similar

73
00:03:34,400 --> 00:03:39,133
to the features of the point
that I'm about to add into the data set.

74
00:03:39,633 --> 00:03:41,933
So basically it falls into that circle.

75
00:03:41,933 --> 00:03:46,200
And so p of x is the number of similar
observations or similar observations

76
00:03:46,200 --> 00:03:49,866
means observations similar to the points
that we're about to add

77
00:03:50,266 --> 00:03:53,933
divided by the total number
of observations, which is student okay.

78
00:03:54,300 --> 00:03:55,566
So it's four four.

79
00:03:55,566 --> 00:03:59,166
We can see that there's four points
in this circle right now divided by 30.

80
00:03:59,833 --> 00:04:05,400
And the interesting thing here is that
this result is the same both times.

81
00:04:05,400 --> 00:04:06,800
So this is in step two.

82
00:04:06,800 --> 00:04:11,233
In step one where we were calculating for
the people who walk to work.

83
00:04:11,233 --> 00:04:14,466
So the probability for people
who walk to work it was the same.

84
00:04:14,466 --> 00:04:17,866
So basically p of x doesn't change
whether you're calculating it in

85
00:04:18,300 --> 00:04:21,266
in the in step one
where we were calculating the probability

86
00:04:21,266 --> 00:04:24,533
that the person with these features,
as a person who walks to work,

87
00:04:24,533 --> 00:04:28,033
or if you're calculating in the step
two scenario where you're calculating

88
00:04:28,033 --> 00:04:31,133
if that person with these features
is a person who drives to work

89
00:04:31,800 --> 00:04:33,833
and therefore it's the same both times.

90
00:04:33,833 --> 00:04:35,533
So what does that mean?
Well, let's have a look at the formula.

91
00:04:35,533 --> 00:04:39,366
So you can see that the formula in
step one it was p of the probability

92
00:04:39,366 --> 00:04:42,366
of person
who exhibits features x that he walks.

93
00:04:42,366 --> 00:04:44,566
He or she works at walks to work.

94
00:04:44,566 --> 00:04:45,933
And that was a formula.

95
00:04:45,933 --> 00:04:48,000
And CP of X is at the bottom.

96
00:04:48,000 --> 00:04:50,433
And for step two
it was a probability of a person

97
00:04:50,433 --> 00:04:54,900
who exhibits features
X being a person who drives to work.

98
00:04:55,066 --> 00:04:57,100
And as you can see,
p of x is at the bottom here.

99
00:04:57,100 --> 00:05:00,400
So what did we do in step three. Right.

100
00:05:00,400 --> 00:05:02,100
So let's move on to step three from here.

101
00:05:02,100 --> 00:05:03,633
Step three we compared the two.

102
00:05:03,633 --> 00:05:08,000
So now if we take these two formulas
these right sides of the formulas

103
00:05:08,366 --> 00:05:10,400
and put them into the comparison

104
00:05:10,400 --> 00:05:13,533
like that, you will see that at the bottom
the denominator is the same.

105
00:05:13,900 --> 00:05:16,566
Now that we know
that the nominator is not zero

106
00:05:16,566 --> 00:05:20,400
and it's actually greater than zero,
probabilities is never less than zero.

107
00:05:20,400 --> 00:05:23,066
And we know it's not zero.
So we can just get rid of it. Right.

108
00:05:23,066 --> 00:05:27,900
We can multiply both sides by p of x
and therefore the sign won't change.

109
00:05:27,900 --> 00:05:30,400
And also we'll get rid of the denominator.

110
00:05:30,400 --> 00:05:33,866
And that way we won't actually have
to perform that calculation.

111
00:05:33,866 --> 00:05:35,566
So that's one less calculation to perform.

112
00:05:35,566 --> 00:05:39,533
So you can just compare the top parts
of these calculations.

113
00:05:39,833 --> 00:05:42,433
And so that is what is done
a lot of the time.

114
00:05:42,433 --> 00:05:46,200
So if you've done other courses
on machine learning or if you've read some

115
00:05:46,200 --> 00:05:50,500
maybe articles on machine learning,
you will find that this is oftentimes

116
00:05:50,500 --> 00:05:50,933
a case.

117
00:05:50,933 --> 00:05:55,266
And also sometimes it is not mentioned
that this is happening.

118
00:05:55,266 --> 00:06:00,000
So sometimes it's assumed that or it can
be assumed that you know what's going on.

119
00:06:00,000 --> 00:06:02,366
So just be careful that look out for that.

120
00:06:02,366 --> 00:06:04,500
And it's a it's a valid approach.
As we discussed.

121
00:06:04,500 --> 00:06:06,400
It's totally valid to do it that way.

122
00:06:06,400 --> 00:06:09,600
But that is
only if you're comparing the two, right.

123
00:06:09,600 --> 00:06:12,233
If you're only comparing the two,
then you can do that

124
00:06:12,233 --> 00:06:14,500
if you actually
want to calculate the value.

125
00:06:14,500 --> 00:06:16,800
So we said 75%, 25%.

126
00:06:16,800 --> 00:06:20,366
If you want to calculate the value,
you can't do that because it'll be to

127
00:06:20,566 --> 00:06:24,566
different realities because you're
supposed to divide by a certain value.

128
00:06:24,566 --> 00:06:26,000
It's not the actual value.

129
00:06:26,000 --> 00:06:30,766
And moreover,
if you want to like comparing is okay,

130
00:06:30,766 --> 00:06:35,066
but calculating the actual value
and maybe some performing some operations

131
00:06:35,066 --> 00:06:38,833
or, you know, you calculate the value
from this, you want to compare the value

132
00:06:38,833 --> 00:06:42,533
from this scenario
to a value from another problem, right?

133
00:06:42,533 --> 00:06:43,566
Like a value

134
00:06:43,566 --> 00:06:48,366
from a different kind of a scenario
where p of x will be different, right?

135
00:06:48,400 --> 00:06:53,633
Not from this particular example
that you're working with.

136
00:06:53,633 --> 00:06:56,033
You want to compare the probability
of the person being a person

137
00:06:56,033 --> 00:07:00,100
who walks to work in this example,
to the probability of that person

138
00:07:00,300 --> 00:07:03,600
being a person who walks to work
from a different example, right

139
00:07:03,600 --> 00:07:04,966
where p of x will be different.

140
00:07:04,966 --> 00:07:08,133
If you want to compare across like that,
that will also not work,

141
00:07:08,133 --> 00:07:10,766
because if p of x is different,
so be careful of that.

142
00:07:10,766 --> 00:07:14,166
It's always the safer and safer
way is always just to perform the full

143
00:07:14,166 --> 00:07:15,200
calculation.

144
00:07:15,200 --> 00:07:18,600
But if you're doing it often,
or if you just want to save some time

145
00:07:18,933 --> 00:07:20,100
or if you just maybe

146
00:07:20,100 --> 00:07:23,800
reading other literature, then it's it's
also good to know about this approach

147
00:07:23,800 --> 00:07:28,200
where the denominator can be dropped
when just comparing between the two.

148
00:07:28,833 --> 00:07:29,133
All right.

149
00:07:29,133 --> 00:07:33,700
So that was another point of like
kind of a hint there or maybe a shortcut.

150
00:07:34,266 --> 00:07:38,566
And so the last point for today,
what happens when we have more than two

151
00:07:38,566 --> 00:07:39,600
classes.

152
00:07:39,600 --> 00:07:42,800
So as you remember in this scenario
we only had two classes, the red and green

153
00:07:42,800 --> 00:07:45,800
or the people who walk to work
and the people who drive to work.

154
00:07:46,100 --> 00:07:48,200
What happens when you have more classes?

155
00:07:48,200 --> 00:07:50,666
How is the challenge different?

156
00:07:50,666 --> 00:07:54,400
Well, when we have only two classes,
we compared, as you remember,

157
00:07:54,400 --> 00:07:56,066
we compared the people.

158
00:07:56,066 --> 00:07:59,933
The probability that a person
who exhibits features X walks to work.

159
00:07:59,933 --> 00:08:01,966
So basically that new data point
that we added,

160
00:08:01,966 --> 00:08:04,200
what is the probability
that it's a person who walks to work

161
00:08:04,200 --> 00:08:06,266
versus what's the probability
that that new data point

162
00:08:06,266 --> 00:08:10,200
is a person who drives to work
and it turned out that we were comparing

163
00:08:10,433 --> 00:08:15,400
75% versus 25%,
and 75% was greater than 25%.

164
00:08:15,400 --> 00:08:18,633
And therefore the probability
of the person of that person

165
00:08:18,633 --> 00:08:19,033
being a person

166
00:08:19,033 --> 00:08:23,100
who walks to work was greater than that
person being a person who drives to work,

167
00:08:23,100 --> 00:08:26,800
and therefore we decide to classify them
as a person who walks to work.

168
00:08:27,166 --> 00:08:28,966
And it's very straightforward.

169
00:08:28,966 --> 00:08:32,766
And moreover, you'll find that every time
when you just have two classes,

170
00:08:33,133 --> 00:08:34,600
it will always add up to one.

171
00:08:34,600 --> 00:08:37,600
So we didn't even have to calculate
the second one.

172
00:08:37,600 --> 00:08:39,333
We could have just stopped at this one,

173
00:08:39,333 --> 00:08:42,933
because if this is 75%,
this one is automatically 25%.

174
00:08:43,200 --> 00:08:44,266
It's always going to be like that.

175
00:08:44,266 --> 00:08:45,933
If you have two classes.

176
00:08:45,933 --> 00:08:49,100
The way it changes, if you have three
classes, it is it gets more interesting.

177
00:08:49,100 --> 00:08:51,200
Right?
So you've calculated one and there's this.

178
00:08:51,200 --> 00:08:52,766
Then there's still two more left.

179
00:08:52,766 --> 00:08:56,800
So if you're working with one
with just two classes,

180
00:08:56,800 --> 00:08:58,366
as soon as you calculate one you're done.

181
00:08:58,366 --> 00:09:01,366
You can decide right away
if it's greater than 50%.

182
00:09:01,366 --> 00:09:03,466
Then you assign that class.

183
00:09:03,466 --> 00:09:05,333
If it's less than 50%, yes, another class.

184
00:09:05,333 --> 00:09:08,366
If it's equal to 50%,
then you've got like a tie.

185
00:09:09,000 --> 00:09:12,266
Whereas if you have 2 or 3
or more classes, then just calculating

186
00:09:12,266 --> 00:09:15,066
the one won't be enough,
because then you have two other ones

187
00:09:15,066 --> 00:09:17,633
and you would still have to calculate
at least another one.

188
00:09:17,633 --> 00:09:21,300
So just means that it's
more of an interesting

189
00:09:21,500 --> 00:09:24,866
selection problem
when you have more classes.

190
00:09:24,866 --> 00:09:29,266
That's pretty much the main thing
that, changes when you have more classes.

191
00:09:29,900 --> 00:09:31,066
And there we go.

192
00:09:31,066 --> 00:09:33,233
That's all for today.

193
00:09:33,233 --> 00:09:37,133
I hope you enjoyed these extra
tips on the naive Bayes classifier,

194
00:09:37,466 --> 00:09:39,333
and I look forward to see you next time.

195
00:09:39,333 --> 00:09:41,166
Until then, enjoy machine learning.
