1
00:00:03,630 --> 00:00:06,840
So far we talked about classification
in terms of just predicting,

2
00:00:06,840 --> 00:00:11,744
is this a positive sentence or a negative
sentence, is email spam or not spam,

3
00:00:11,744 --> 00:00:15,210
but in general,
you wanna go a little bit beyond that and

4
00:00:15,210 --> 00:00:19,550
ask about, what is the probability
that this is email spam?

5
00:00:19,550 --> 00:00:22,500
So how confident am I on the prediction?

6
00:00:22,500 --> 00:00:24,260
So if you're just looking at positive and
negative,

7
00:00:24,260 --> 00:00:27,200
I want to know how sure I
am about the prediction.

8
00:00:27,200 --> 00:00:30,550
So for example,
if you take a sentence like, the sushi and

9
00:00:30,550 --> 00:00:32,110
everything else were awesome.

10
00:00:33,540 --> 00:00:35,920
That's a definite plus.

11
00:00:35,920 --> 00:00:37,730
Definitely positive.

12
00:00:37,730 --> 00:00:41,330
However, the sushi was good,
the service was okay.

13
00:00:41,330 --> 00:00:43,890
It's probably a plus but I'm not so sure.

14
00:00:43,890 --> 00:00:45,810
It's not as definite.

15
00:00:45,810 --> 00:00:51,824
And so what a classifier will often do,
is not just output is positive or

16
00:00:51,824 --> 00:00:56,800
negative, but output how confident,
how sure it is.

17
00:00:56,800 --> 00:00:59,440
One way to do that is to
talk about probabilities.

18
00:00:59,440 --> 00:01:03,352
So you have to play the probability
of being a positive or

19
00:01:03,352 --> 00:01:07,570
negative sentence,
given the input sentence x.

20
00:01:07,570 --> 00:01:08,810
So the output label,

21
00:01:08,810 --> 00:01:12,190
what's the probability output label,
given the input sentence?

22
00:01:12,190 --> 00:01:19,219
So for example, for the top example there,
instead of saying that's definite +,

23
00:01:19,219 --> 00:01:25,530
we say the probability that
it's a + given x in is 0.99.

24
00:01:25,530 --> 00:01:29,880
Even x is only 0.55 because I'm
kind of uncertain about that.

25
00:01:29,880 --> 00:01:34,149
Predicting probabilities or level of
confidence is extremely important and

26
00:01:34,149 --> 00:01:38,430
as we'll see in the classification course,
it allows you to do many things.

27
00:01:38,430 --> 00:01:42,786
So for example when you know the
probability, you can make decisions like,

28
00:01:42,786 --> 00:01:46,667
what is a good decision boundary
that trades off false positives and

29
00:01:46,667 --> 00:01:49,480
false negatives, and
balance between the two.

30
00:01:49,480 --> 00:01:53,749
[MUSIC]