1
00:00:00,020 --> 00:00:04,668
[MUSIC]

2
00:00:04,668 --> 00:00:10,151
Now that explored the idea of precision
let's talk about the slightly more subtle

3
00:00:10,151 --> 00:00:15,237
idea of recall, and recall has to do
with out of the possible positive things

4
00:00:15,237 --> 00:00:20,720
in the world, everything out there,
did I find every possible positive?

5
00:00:20,720 --> 00:00:22,970
Sentence about my restaurant.

6
00:00:22,970 --> 00:00:27,730
So, more specifically, I might take
all the restaurant reviews and

7
00:00:27,730 --> 00:00:31,290
all the sentences and
feed it through my classifier.

8
00:00:31,290 --> 00:00:35,270
And then classify some of those as
positive, and some of those as negative.

9
00:00:35,270 --> 00:00:37,540
So, some might have, y hat = +1.

10
00:00:37,540 --> 00:00:39,530
Some might have y hat = -1.

11
00:00:39,530 --> 00:00:45,342
Now, if I look at the true label, the true
class of each one of the sentences, so yi,

12
00:00:45,342 --> 00:00:50,870
which ones have +1, you see that four
of those were predicted to be positive,

13
00:00:50,870 --> 00:00:53,190
and just like we talked
earlier in the module.

14
00:00:53,190 --> 00:00:58,230
But there turned out to be two other
sentences, there were three positive

15
00:00:58,230 --> 00:01:02,270
that my algorithm did not find
that fell into the negative bin.

16
00:01:02,270 --> 00:01:04,300
So those are missed out.

17
00:01:04,300 --> 00:01:09,030
So in other words,
we found the 4 positive sentences, but

18
00:01:09,030 --> 00:01:14,100
we missed out on 2 positive sentences
that could have shown in my website.

19
00:01:14,100 --> 00:01:15,730
And maybe those two sentences were so

20
00:01:15,730 --> 00:01:22,230
amazing they would change the history
of my restaurant, but I missed it.

21
00:01:22,230 --> 00:01:25,100
Because they did not have perfect recall.

22
00:01:25,100 --> 00:01:29,890
So, more formally out of all
the truly positive data points, so

23
00:01:29,890 --> 00:01:36,190
the one's where yi is +1, we can have
a subset on those that we do capture.

24
00:01:36,190 --> 00:01:41,040
So where y hat is also +1 but
there's a subset which we

25
00:01:41,040 --> 00:01:46,620
don't capture where we think the y hat
is -1, so y hat does not agree with y.

26
00:01:46,620 --> 00:01:48,450
And so, that's the part that we missed.

27
00:01:48,450 --> 00:01:52,850
So the recall is the fraction of
the once that we actually get.

28
00:01:52,850 --> 00:01:54,520
We want everything to be
in the blue box here.

29
00:01:55,770 --> 00:01:59,650
More formally, we can define recall
as the fraction of true positives.

30
00:01:59,650 --> 00:02:02,340
For these are the data points
that we were positive and

31
00:02:02,340 --> 00:02:05,370
we got them right divided
by the true positives.

32
00:02:06,410 --> 00:02:07,580
And the false negative.

33
00:02:07,580 --> 00:02:11,550
So the data points that were true,
but we labeled as negative.

34
00:02:11,550 --> 00:02:13,700
So falsely labeled as negative.

35
00:02:13,700 --> 00:02:21,940
And so, this is going to have value
one if the false negatives are zero.

36
00:02:21,940 --> 00:02:26,250
Which means we captured everything,
we captured all the true positives.

37
00:02:26,250 --> 00:02:30,530
And zero if we did not
capture any true positives so

38
00:02:30,530 --> 00:02:33,570
all the positive data went
to the false negative bin.

39
00:02:34,690 --> 00:02:38,620
So if we go back to the example
that we have been looking at,

40
00:02:38,620 --> 00:02:42,490
I want to show positive
sentences on my website.

41
00:02:42,490 --> 00:02:50,400
I've got four of them, in y hat i equals
+1 but I missed out on two sentences.

42
00:02:50,400 --> 00:02:51,110
So for example,

43
00:02:51,110 --> 00:02:57,270
I missed out on the sentences that said my
wife tried the ramen and it was delicious,

44
00:02:57,270 --> 00:03:01,460
and so maybe somebody's interested in
ramen they don't see that sentence,

45
00:03:01,460 --> 00:03:05,853
they don't go to my website, and so
I missed out on something really good.

46
00:03:05,853 --> 00:03:10,489
So high recall means that you discover
everything positive that's being

47
00:03:10,489 --> 00:03:14,388
said about the restaurant or
all of the positive data points.

48
00:03:14,388 --> 00:03:19,669
[MUSIC]