1
00:00:00,000 --> 00:00:06,009
When we pick up a newspaper, we are
looking for the maximum information

2
00:00:06,009 --> 00:00:10,031
content.
So, naturally, more surprising events make

3
00:00:10,031 --> 00:00:15,012
for better news.
And in passing, as you read the news, you

4
00:00:15,012 --> 00:00:20,087
and I glance at some advertisements, and
the paper makes some money.

5
00:00:21,062 --> 00:00:29,058
Of course, the questions the newspaper
needs to ask is, when to place an ad, and

6
00:00:29,058 --> 00:00:35,079
where to place an ad?
Which newspapers against which stories?

7
00:00:35,079 --> 00:00:41,850
Unfortunately print newspapers just don't
have that freedom.

8
00:00:41,850 --> 00:00:49,332
Sometimes the interesting news is on the
sports page, whereas your most expensive

9
00:00:49,332 --> 00:00:54,496
ad is right up front.
To understand how better advertising

10
00:00:54,496 --> 00:01:02,210
model, in particular online advertising,
is related to information theory, let's

11
00:01:02,210 --> 00:01:09,323
first see what Shannon did with his
concept of information being related to

12
00:01:09,323 --> 00:01:14,671
surprise.
Shannon was concerned with communication

13
00:01:14,671 --> 00:01:22,287
along a noisy channel, like a telephone
line or the radio waves on the air.

14
00:01:22,287 --> 00:01:27,530
Today he'd be concerned about
communicating on the internet.

15
00:01:27,530 --> 00:01:32,358
But of course the internet wasn't there in
his times.

16
00:01:32,358 --> 00:01:41,323
The model that he used was the transmitted
signal at one end of the channel.

17
00:01:41,323 --> 00:01:49,645
It could be a sequence of messages or,
today, a sequence of bits received on the

18
00:01:49,645 --> 00:01:54,959
other side as another sequence of
messages.

19
00:01:54,959 --> 00:02:02,993
There's an important concept that he
defined called the mutual information

20
00:02:02,993 --> 00:02:07,427
between the transmitted signal and the
received signal.

21
00:02:07,427 --> 00:02:14,497
And the purpose of the channel, these days
the cell phone network or the internet is

22
00:02:14,497 --> 00:02:21,494
to maximize the mutual information between
what is transmitted and what is received.

23
00:02:21,494 --> 00:02:28,345
So obviously, you really want to hear what
your friend is saying at the other end of

24
00:02:28,345 --> 00:02:34,316
the line.
Now let's model the advertising problem in

25
00:02:34,316 --> 00:02:39,899
the language of communication along a
noisy channel.

26
00:02:39,899 --> 00:02:49,093
The advertising context we approach, the
newspaper or webpages with some intent.

27
00:02:49,093 --> 00:02:58,050
And some degree of attention and interest.
And the advertiser would like to measure

28
00:02:59,006 --> 00:03:05,036
which pages we are interested in and when
we have an intention to buy.

29
00:03:05,074 --> 00:03:14,039
Unfortunately for normal print or media
advertising, the only signal available to

30
00:03:14,039 --> 00:03:22,030
an advertiser is the actual sales
transactions at the end of a quarter or

31
00:03:22,030 --> 00:03:30,736
the ad revenue that one spends and after
the fact, correlation of the two is all

32
00:03:30,736 --> 00:03:42,180
that one is really able to do.
This is highly erroneous and so the mutual

33
00:03:42,180 --> 00:03:51,129
information between our actual intent and
attention and the lag measures, which come

34
00:03:51,129 --> 00:04:00,487
much later of sales transactions or
advertising revenue incurred is very low.

35
00:04:00,487 --> 00:04:09,292
In other words, it's very difficult to
predict which paper to advertise in.

36
00:04:09,292 --> 00:04:16,336
And on which page do advertising simply
based on this signals.

37
00:04:16,336 --> 00:04:20,970
The online world of course changes
everything.

38
00:04:20,970 --> 00:04:27,348
Because the signals we have are the clicks
made by users, the queries made when they

39
00:04:27,348 --> 00:04:31,887
search, and the actual content that they
are reading.

40
00:04:31,887 --> 00:04:36,745
The mutual information therefore is much
higher.

41
00:04:36,745 --> 00:04:43,517
Let's see how.
A brief overview of online advertising

42
00:04:43,517 --> 00:04:48,219
first.
In the case of search advertising

43
00:04:48,219 --> 00:04:55,378
advertisers bid for keywords in Google's
online keyword option, which is happening

44
00:04:55,378 --> 00:05:02,021
continuously all the time.
The highest bidders ads are place against

45
00:05:02,021 --> 00:05:08,076
the matching searches, so if y bid for
shoes and you searched for shoes and mine

46
00:05:08,076 --> 00:05:14,082
was the highest bid, then my ad would come
up against your query and so on.

47
00:05:14,082 --> 00:05:21,053
The actual auction mechanism is a bit more
complex than that, but we won't go into

48
00:05:21,053 --> 00:05:27,090
that right now.
The result of this advertising model is

49
00:05:27,090 --> 00:05:35,047
that the mutual information between the
advertising dollars, spent, and the sales

50
00:05:35,047 --> 00:05:42,075
that they translate into is improved
considerably from traditional media, and

51
00:05:42,075 --> 00:05:49,010
is fundamentally responsible for the
success of online advertising.