1
00:00:09,580 --> 00:00:13,280
Recommendation systems are
used in many different areas

2
00:00:13,280 --> 00:00:15,150
other than movies.

3
00:00:15,150 --> 00:00:18,960
Jeff Bezos, the
CEO of Amazon, said

4
00:00:18,960 --> 00:00:22,110
that, "If I have 3 million
customers on the web,

5
00:00:22,110 --> 00:00:25,430
I should have 3 million
stores on the web."

6
00:00:25,430 --> 00:00:28,700
The internet allows for
mass personalization,

7
00:00:28,700 --> 00:00:32,439
and recommendation systems
are a key part of that.

8
00:00:32,439 --> 00:00:34,560
Recommendation
systems build models

9
00:00:34,560 --> 00:00:37,990
about users' preferences
to personalize the user

10
00:00:37,990 --> 00:00:39,620
experience.

11
00:00:39,620 --> 00:00:43,340
This helps users find items they
might not have searched for,

12
00:00:43,340 --> 00:00:45,990
like a new favorite
band, an old friend who

13
00:00:45,990 --> 00:00:49,740
uses the same social media
network, or a book or song

14
00:00:49,740 --> 00:00:52,730
that they're likely to enjoy.

15
00:00:52,730 --> 00:00:55,260
Recommendation systems
are a cornerstone

16
00:00:55,260 --> 00:00:57,450
of these top businesses.

17
00:00:57,450 --> 00:01:00,170
Social networking
sites, like Facebook,

18
00:01:00,170 --> 00:01:04,150
music streaming sites, like
Pandora, and retail companies,

19
00:01:04,150 --> 00:01:07,610
like Amazon, all provide
recommendation systems

20
00:01:07,610 --> 00:01:09,860
for their users.

21
00:01:09,860 --> 00:01:12,840
Both collaborative filtering
and content filtering

22
00:01:12,840 --> 00:01:14,740
are used in practice.

23
00:01:14,740 --> 00:01:17,230
Collaborative filtering
is used by companies

24
00:01:17,230 --> 00:01:20,850
like Amazon, Facebook,
and Google News.

25
00:01:20,850 --> 00:01:23,080
Content filtering
is used by companies

26
00:01:23,080 --> 00:01:27,380
like Pandora, Rotten
Tomatoes, and See This Next.

27
00:01:27,380 --> 00:01:30,420
And Netflix uses both
collaborative filtering

28
00:01:30,420 --> 00:01:33,890
and content filtering.

29
00:01:33,890 --> 00:01:37,120
So now let's go back
to the Netflix prize.

30
00:01:37,120 --> 00:01:42,789
29 days after last call was
announced, on July 25, 2009,

31
00:01:42,789 --> 00:01:49,020
the team The Ensemble submitted
a 10.09% improvement, beating

32
00:01:49,020 --> 00:01:52,700
the 10.05% improvement that
was submitted by Bellkor's

33
00:01:52,700 --> 00:01:56,190
Pragmatic Chaos to
signal last call.

34
00:01:56,190 --> 00:01:59,680
But by the time Netflix stopped
accepting submissions the next

35
00:01:59,680 --> 00:02:04,830
day, Bellkor's Pragmatic Chaos
had also submitted a 10.09%

36
00:02:04,830 --> 00:02:08,930
improvement, and The Ensemble
had submitted a 10.10%

37
00:02:08,930 --> 00:02:10,620
improvement.

38
00:02:10,620 --> 00:02:13,340
To really test the
algorithms, Netflix

39
00:02:13,340 --> 00:02:15,470
tested them on a
private test set

40
00:02:15,470 --> 00:02:17,800
that the teams had
never seen before.

41
00:02:17,800 --> 00:02:21,840
This is the true test
of predictive ability.

42
00:02:21,840 --> 00:02:26,480
On September 18, 2009, Netflix
announced that the winning team

43
00:02:26,480 --> 00:02:29,200
was Bellkor's Pragmatic Chaos.

44
00:02:29,200 --> 00:02:34,490
They won the competition and
the $1 million grand prize.

45
00:02:34,490 --> 00:02:37,470
Recommendation systems
provide a significant edge

46
00:02:37,470 --> 00:02:39,310
to many companies.

47
00:02:39,310 --> 00:02:42,860
In today's digital age,
businesses often have hundreds

48
00:02:42,860 --> 00:02:45,980
of thousands of items to
offer their customers,

49
00:02:45,980 --> 00:02:49,640
whether they're movies, songs
, or people they might know

50
00:02:49,640 --> 00:02:51,340
on Facebook.

51
00:02:51,340 --> 00:02:54,160
Excellent recommendation
systems can make or break

52
00:02:54,160 --> 00:02:56,140
these businesses.

53
00:02:56,140 --> 00:02:58,430
Clustering algorithms,
which are tailored

54
00:02:58,430 --> 00:03:01,970
to find similar customers
or similar items,

55
00:03:01,970 --> 00:03:06,030
form the backbone of many of
these recommendation systems.

56
00:03:06,030 --> 00:03:10,050
Clustering also has many other
interesting applications.

57
00:03:10,050 --> 00:03:12,720
In the next lecture,
we'll see how clustering

58
00:03:12,720 --> 00:03:16,060
can be used to improve
the predictive ability

59
00:03:16,060 --> 00:03:18,650
of classification methods.