Can Facebook feed its ad brains?

By Stefanie Olsen
http://www.news.com/Can-Facebook-feed-its-ad-brains/2100-1024_3-6216553.html 

November 2, 2007 

Facebook is now valued at $15 billion by many estimates. Soon the 
young company will have to prove it's really worth it. Next week in 
New York, Facebook is expected to unveil its new advertising strategy 
on the heels of an expanded advertising partnership and $240 million 
investment from Microsoft. The company will likely talk about selling 
targeted ads to members, but experts say that to build its own ad 
serving system, Facebook must figure out how to serve the right ads 
to the right people in real time.

"That's a very difficult problem at large scale, with so many ads and 
millions of people," said Greg Linden, founder of Findory and an 
architect of Amazon.com's early product recommendation and 
personalization engine. "And the data's not well tied to purchases 
like at Amazon, or even like Google, where so many Web searches 
are about products."

The predominant question on everyone's mind is: Can Facebook 
build an ad system clever enough to keep pace with the passing 
fancies of its social-networking members? Facebook will have to get 
really good at processing all of the data it has collected on its 
reported 49 million members--demographics, personal preferences, 
and social histories--to predict what advertisements they might 
actually like and respond in their "news feed" or next to their "wall," 
according to industry executives.

That's no small task. In fact, it's a massive computing problem and 
one that very few companies apart from Google and Amazon have 
mastered. That's why Facebook--a company known for its young, 
fun culture--has been trying to hire more seasoned experts in 
so-called machine learning who can develop the right algorithms for 
a new generation of targeted advertising, people familiar with the 
company say.

One tech executive characterized the challenge like this: "The 
company that can process the most data will win." That maxim has 
proven true of Google in Web search and Amazon in e-commerce 
sales and product recommendations. Now Facebook must figure out 
how to take billions of data points about its members and turn that 
into an automatic ad machine. 

A Facebook representative declined comment for the story.

There's no question Facebook is sitting on a data goldmine, with an 
exhaustive amount of information on people's preferences, backgrounds, 
and social histories--all given voluntarily by members. Facebook has 
profiles that include people's favorite music, television shows, books, 
and hobbies; their job history, education, birth date, and marital 
status; as well as daily activities, social networks, and interest 
groups. Traditional ad networks would kill for all that information in 
one place.

But with that data comes some interesting machine learning problems, 
experts say.

"The problem isn't that they can't make revenue, it's that the 
expectation is so high on the amount of revenue that they can make." 
--Greg Linden, Findory founder Machine learning is a broad term in 
the field of artificial intelligence. It refers to developing algorithms that 
can discover patterns in data and learn from them. Google, for example, 
has used probabilistic Bayesian models to serve results to data 
searches based on keywords. With advertising, it's all about matching 
the right person to the right ad. And on an individual level, that's a 
tall order.

So some technologists focus on lumping people into groups or types, 
tracking their typical behaviors in aggregate, and then trying to 
predict what they might want or do next.

Machine learning in online advertising might involve trying many 
different techniques on affinity groups to figure out which work best. 
That's because no one obvious technique is the silver bullet for 
social networks--no one has solved the problem of serving ads in 
that setting before.

Many thorny issues can arise, too, such as trusting what people 
post about themselves. Social networks can be noisy in terms of 
data, meaning that one day a member can say he broke it off with 
a girlfriend and change everything about his music and film tastes. 
On social networks, people are prone to misspellings, random 
statements, and exaggeration. Also, Web surfers are on Facebook 
to socialize, network, or be entertained, not to buy something. 
In that kind of social environment, ads can be ineffective or annoying.

That's why Facebook must perfect a subtle product placement or 
recommendation system. To do it, it will have to invent algorithmic 
tricks. For example, knowing a list of people's friends isn't necessarily 
useful unless the system could automatically remind people of 
birthdays, and then advertise a specific gift the friend might like 
based on his or her preferences.

Aggregate Knowledge in Palo Alto, Calif., which is backed by 
Google investor Kleiner Perkins Caufield & Byers, may also have 
an ad solution for social networks. Aggregate has developed 
algorithms to determine what are called "affinity clusters" of people 
and, based on the personality profiles of those people, targets ads. 
It does this by looking at people's habits in aggregate, rather than 
as individuals.

"We do a contest of algorithms in each context...and see which 
works best (to serve an ad) and that's a traditional machine learning 
technique. But this requires massive computing power," said 
Paul Martino, founder of Aggregate Knowledge. Martino said his 
company is in talks with various social networks, but does not yet 
have a deal with any of them.

One of the techniques in this field is known as collaborative filtering, 
which Amazon used when creating its product recommendation 
system. Amazon's system automatically analyzes your purchase 
history and looks for the same buying patterns among other 
shoppers. By sizing up the purchase histories of similar shoppers, 
the system can look for the products you haven't bought yet and 
that other similar shoppers have. Then it can suggest items you 
might also like.

Facebook plans to adopt similar techniques, for people who like the 
same music or films, according to people familiar with the company. 
That way, movie or music studios could "suggest" entertainment in 
the form of a product placement.

Behavioral targeting

Another approach is to tailor ads to a person's demonstrated behaviors, 
or what's called behavioral targeted advertising. That means that a 
site might keep track of a person over time and factor in his or her 
demographics and preferences. A high-income woman who has 
recently said she's looking for a car might receive a Lexus ad, 
for example.

Facebook has already come up with some machine learning tricks 
to allow people to search on a person's nickname, even if the 
person hasn't divulged that information. According to one source, 
the company developed an algorithm that could strip words from 
people's "wall" (where friends post messages) and then remove 
from the list all the words that weren't in the dictionary. With what 
remained it built a comprehensive dictionary of nicknames so that 
people can search on a friend's profile under alternative names.

"That's an interesting machine learning problem, and they have a 
million things like that," said the source.

Facebook is already clustering people into groups, such as tech 
geeks or music lovers, and following patterns of people's behaviors 
to predict other kinds of behaviors, according to people familiar with 
the company. For example, groups who like baseball could like 
sushi, in a hypothetical link within data patterns that you might not 
expect. The ad network could then target ads for local sushi 
restaurants to members of that group. But the company hasn't 
deployed this methodology on a wide scale.

Facebook is also a tremendous barometer for public opinion. 
So the site could eventually track chatter about a movie like 
Spider-Man, for example, and sell that information to the film 
company or watch how a message about the movie on a wall is 
received. If the chatter falls off quickly, it could be time for the 
studio to release the movie on DVD rather than keep it in theaters.

Microsoft serves ads based on audience segment, such as car 
shoppers, by anonymously tracking the behavior of users across 
its network. If an MSN user has browsed MSN Autos or searched 
for "Kelly Blue Book" on Live Search, the user will see relevant ads. 
The search and Web browsing history data is blended with data 
people offer to Microsoft during registration, such as age and gender, 
but not identifiable information.

For now, Microsoft will be serving Facebook's graphical ads, but it 
remains to be seen how the social network will tackle the ad issue 
on its own.

"The problem isn't that they can't make revenue, it's that the expectation 
is so high on the amount of revenue that they can make," Linden said. 
"Because people aren't in a purchasing frame of mind (at Facebook), 
it's going to be hard for them to get as much from the advertising as 
the hype right now."