biographies story podcast

The Empathetic Internet

Thanks to new developments in data mining, computers are learning increasingly detailed profiles of individual users. Erin Digitale wonders how well she wants the web to get to know her. Illustrated by Kacie Carter.

Illustration: Kacie Carter

There it is, right on top of the recommended movie list in my familys Netflix account: Mr. Bean's Holiday.

Ugh, I think. I hate Mr. Bean. Has Netflix lost its cookies?

No, but it is having trouble making good use of its vast stockpile of user-generated movie ratings. Thats why the popular web-based DVD rental company is offering a $1 million prize to improve its movie-recommendation software. One contestant is UC Santa Cruz computer scientist Yi Zhang.

Zhang and her team study how computers learn from the data we generate when, for example, we search the Internet or make online purchases. Were trying to come up with algorithms to train machines to think like human beings, says Anita Krishnakumar, a graduate student working with Zhang.

The Netflix contest is just one part of Zhangs overall efforts. Her teams research, which trains computers to give tailored responses to individual users, will hone search-engine results and help businesses target online advertising. The research will improve your computers ability not just to fetch information from queries, but also to send you the content you needunprompted. Ultimately, it aims to transform the consumer portion of the Internet from an unwieldy behemoth with the simple stimulus-response intelligence of a lizard to a well-trained, useful beast with the ability to sort information rationally.

Most of Zhang's peers focus on isolated aspects of this field, called data mining, such as exploiting online social networks or helping computers understand language. Zhang says a narrow focus would limit the usefulness of her research, so she teaches computers to integrate many kinds of data. She makes novel use of Bayesian analysis, a branch of statistics that gives computers a logical architecture for sorting and learning from new information. The human mindour headis a unifying framework, and Im trying to build that, she says.

But for all its potential benefits, data-mining research like Zhangs also raises serious questions about computer security and personal privacy. The fact that databases are learning to talk to each other presents a distinct set of challenges for our legal system, says Lauren Gelman, director of the Center for Internet and Society at Stanford Law School. Whereas in the past, tracking an individual through public records required digging through file drawers in courthouses, today the same information is available with a few keystrokes. Now, we have a very different privacy question, but the laws have not changed, says Gelman.

Measuring our longings

Selective data filtering is nothing new. Mail-order specialists at Sears and Montgomery Ward were knee-deep in customer data by the turn of the 20th century. The New York-based Direct Marketing Association was founded in 1917 to help make sense of what direct-mail customers wanted. And clipping services, which let organizations track media reports on their activities, have been going strong for about 100 years.

In the old days, picking out trends in these records was a fuzzy business. For instance, mail-order companies could track consumers names, addresses, and some demographic tidbits such as gender or age. But fine details on customers habits were inaccessible. A turn-of-the-century clothing merchant had no way to figure out which young ladies lingered covetously over his advertisement for an Inflatable Elastic Bosomthe one that adjusted for all four stages of a womans life (Miss, Debutante, Mother, and Dowager) and doubled as a life preserver.

Today, its different. When I search inflatable bosom on, I get three books on the history of womens fashions. Using my computers Internet Protocol (IP) address, Amazon can track which books I click on, how long I spend reading the product descriptions, whether I search the books text, and what I decide to buy.

The key to such intense data collection is surplus computing power. Its now easy to trail users all over the Internet in minute detail, and cheap to store the maps of their trails. And because the torrents of data seem like they might be useful someday, companies are reluctant to ditch even the most mundane tidbits.

Climbing Mt. Data

But accumulating all that detail causes a new problem: mountains of numbers. Its not just using the Internet that leaves dense data trails. Scientists produce stacks of data with high-powered telescopes that scan the skies, genetic investigations that sort millions of DNA codings, and public-health studies that track scores of people over many decades. On a personal scale, all cellphone calls generate data. So do bank transactions, visits to the doctor, and trips to the grocery store.

Yi Zhang is one of many researchers tunneling into this mountain. She wants to enable computers to judge the data in a rational way, as a person would. She knows this goal is a long way off. Meanwhile, we are making things that are useful for people, she says.

Zhang works from a shiny green-glass building that looks like a silicon computer chip dropped among the redwood trees of UC Santa Cruz. Her office is sleek: modular desk, low couch, few books. It seems too emptyuntil she opens a drawer and pulls out stacks of CDs from Netflix, Microsoft, PubMed, and Reuters.

I like to play with this real data, she says.

Zhang didnt always see data as the raw material for play. At first, sifting through numbers struck her as simple, even boring. But as she tried to make computers mimic human minds, the data increasingly drew her in: I finally made what seems like a boring problem into a very interesting problem. She enjoys slicing smart computing into manageable bites. Right now, she says, although we cannot build a robot that can talk intelligently, we can build something in the middle, like a search engine.

Search engines such as Google and Yahoo already use basic data-mining techniques to tailor results to the search histories of users. Suppose you enter spears into a search engine, says Krishnakumar. The first time, it just recommends everything, she says. Specialty websites about medieval spears, a Wikipedia article about spears, Britneys latest anticstheyll all pop up. But read about Britneys legal woes, then do the same search a few more times, and the results change. After three or four days, when you search again, you can see Britney Spears being ranked higher up, Krishnakumar says.

(For the record, I had limited success with this experiment. Everything came up Britney the first time. Of course, the results may have been biased by the fact that my computer knows my secret addiction to the website of People, which pantingly reports Britneys every change of wig.)

Still, todays search engines can only get so close to you. Zhang wants to go further, so her team is building recommendation systems. Search engines make users hunt for new content; recommendation systems suggest it automatically. Netflix, for instance, tracks the films you rent, solicits your ratings, and considers which movies your friends liked. It combines these factors to generate a profile that drives the Films youll heart page. You can view it as a proactive search engine, Zhang says.

Recommendation systems must be taught to question users intelligently, Zhang says, in the same way new friends get to know each other. Rather than asking a rote set of questions, the system has to learn actively and adjust its questions on the fly. That means taking risks. So instead of just suggesting content that matches your profile, recommendation systems also add exploratory suggestions that push the boundaries of your preferences. If you like one of these new items, it opens a big box of treasures for you, Zhang says. Even when you dislike a new item, the system still learns about you. If you spend only a short time looking at a suggestion, or fail to click web links from it, the system makes more inferences about what you like. The systems goals parallel business goals, Zhang notes: I want to make you happy now, and make you happy in the future. 

Like humans getting to know each other, recommendation systems need to be a bit cautious. The information the machine has collected is very unreliable, noisy, and biased, Zhang says. Some people try to cheat the machine, she adds, such as those trying to promote their product.

The system also must withstand the enormous volume of data it encounters. The machine is living in a world thats probably even more complicated than our human world, she says. Once a real system is launched, it will face a lot of peoplemore people than a human faces in his whole life, every single day. Rather than using the brute-force approach of supplying more computing power, Zhangs team is helping computers navigate the data onslaught by showing them which calculations to focus on. Its similar to the filtering that happens in the part of the human brain that unconsciously sorts distracting noises from important sounds.

Building the gears

So computers running recommendation systems must be adventurous, cautious, and adept with cascades of information. How do they translate a bunch of ones and zeroes to these complex traits?

Creating the mathematical gears of a recommendation system requires several steps. Zhangs team first cleans its data by removing points that dont make sense, such as a reported human age of 263 years.

Then they weed. The data have hundreds of different attributes, but it doesnt make sense to use everything. A recommendation system gauging the interests of users might pay attention to their ages and genders but ignore their heights and eye colors.

Once the raw numbers are ready, Zhangs team must make an opening guess at the mathematical equations that will best represent their data. Graduate student Ethan Zhang (no relation) says the teams guess-and-check approach parallels the way other scientists design experiments without the Petri dishes or Erlenmeyer flasks. In the beginning, there are no plans, just trial and error, he says. As the experiments progress, the researchers check their new strategies against older mathematical approaches to see if theyre improving on the old way of doing things.

Zhangs team builds its guesses with the tools of Bayesian analysis, a branch of statistics that combines old information with new data in a rational way. Bayesian analysis translates starting beliefs about the data into mathematical probabilities. The nascent recommendation system begins by treating these probabilities as weights, then asks users for additional weighted data: opinions. If you give a rating of four for movie A, three for movie B, and five for movie C, you are giving a weight to each of the movies, Krishnakumar says. The weighted values add together to give a value for the entire equation. With that value, you can come up with the next item to recommend to the user, she says.

A recommendation system uses your feedback to weight categories of items. The system constructs equations to represent your preferences for movie genres, directors, items your friends liked, or items that got high ratings from people the system thinks are similar to you. With each new tidbit of data, the computer reruns the equations that describe you. The calculations are much simpler than the wiring of your brain, but the deluge of data helps the computer compensate for this simplicity. With each new turn of the equations' cogs, the system learns a little more about you.

Many computer scientists use groups of endlessly recalculated equations to discern your preferences. Zhangs work stands out because it extends the basic idea of reiterated calculations in both breadth and depth. In addition to bringing together many types of data under a unifying framework, Zhangs approach also teaches computers to understand deeper nuances of a users preferences and needs. Its not so hard to develop a model of what you might be interested in, says Jamie Callan, a computer scientist at Carnegie Mellon University who supervised Zhangs Ph.D. research. But its much more difficult to depict the level of detail youre looking for.

This problem becomes significant for recommendation systems that rate pieces of text. Are you reading websites about heart disease because your grandfather just had a heart attack, or because youre a cardiologist checking the newest heart-attack treatments? Those two users, interested in the same topic, need very different things from a recommendation system. The system must gauge the level of detail in the material. Thats a hard problem, Callan says. Many people have struggled with it. Yi was one of the first people to do really good research.

The way Zhang builds her recommendation systems is unique, says Stephen Robertson, a senior researcher at Microsoft who knows her work well. We tend to have a black-box notion of users preferences, Robertson says. Yi tried to disentangle that into different components. Robertson says Zhangs approach is important because, rather than just solving a single narrow search problem in an ad-hoc way, it takes a logical, theory-based approach.

To do this, Zhang built her Bayesian analysis frameworks by consciously considering the four main ways people learn: They use prior knowledge, ask good questions, use context and feedback, and observe the behaviors of others. She selected a different branch of Bayesian analysis to represent each of these types of learning mathematically. Then, she fitted the four analyses into a mathematical frame that shows the computer how the various types of data connect to one another. Getting to know a new friend, you care more about your big similarities (she lives nearby, you enjoy the same activities) than trivial differences (she drinks coffee, you dont). The math frame focuses the computers attention on important similarities among users and helps it ignore the trivia.

Zhangs approach is shedding light on the entire field of data mining, Robertson says. Web search engines are limited in what they can do, he says. Understanding those limitations is where we should be going.

Chasing down our privacy

The increasing power of search engines has a down side, say some observers.

Data is the pollution problem of the information age, says privacy expert Bruce Schneier, who runs the computer security company BT Counterpane in Mountain View, California. The question is whether information gets recycled after users disclose it. The U.S. has no real laws protecting someone from doing something with your information.

The European Unions privacy laws say information collected for one purpose cannot be reused for another. Lack of similar U.S. laws is causing problems. For example, social-networking site Facebook created an uproar in late 2007 with a new feature that automatically broadcast the online activities of users to their friends. Users voiced angry complaints, saying it was too hard to opt out of the broadcast feature. Facebook founder Mark Zuckerberg eventually apologized. People need to be able to explicitly choose what they share, he said in a blog message.

Schneier agrees. But he says most computer users dont understand what theyd have to do to protect their information from being reused. Go ask your mother how to opt out, he says. The rules are hard to find and the technical ability isnt there. And without widespread awareness of the problem, he says, there isnt adequate market demand for businesses to invest in technology that would protect users privacy. Legislation is the only way out, Schneier says.

Lauren Gelman, the Stanford law professor, worries that without enhanced privacy laws, data mining will erode privacy to the point that it dampens our culture. I take an approach to privacy that looks at it as a community interest, a societal interest, she says. If we live in a world where everything about what we do is being captured and catalogued and becomes searchable, then well start thinking twice about what we do. We need to leave holes in the system so people can experiment. Gelmans law students already worry about exchanging experimental ideas in class blogs. What if someone uses those blogs 20 years down the road to derail a Supreme Court nomination? As a society, its to our detriment if we start to self-censor, she says.

Computer scientists, meanwhile, have mixed reactions to the question of how to protect users privacy.

Jamie Callan, Yi Zhangs Ph.D. supervisor, agrees with Gelman and Schneier that the U.S. needs new privacy laws. Its not instantly clear what those laws ought to be, he says, but right now theres very little protection. As work like Zhangs enables computers to integrate many kinds of data, it will become increasingly possible to build incredibly detailed models of what you do every minute of every day, Callan says. And I think thats something the American public ought to be concerned about.

Zhang says shes no expert on privacy law, but she sees promise in technical strategies for protecting privacy. You can design the system so that, for non-public information, you just ask users when they want to share it. Businesses eventually will realize its in their economic interest to make users happy by protecting their privacy, she says. Until that happens, she thinks users should exercise caution about what they put on the Internet.

In the end, data-mining research is double-edged. If researchers manage it correctly, this work will turn the vast, unwieldy Internet into millions of individually tailored internets which feed each of us information suited to our idiosyncratic needs. At that point, Netflix will only recommend Mr. Bean to those who truly appreciate his irritating wackiness.

But to benefit from the many internets of the future, consumers must demand technologies that guard their data or ask for laws prohibiting data reuse. Without these protections, todays Bean-sized nuisance could morph into tomorrows intrusive surveillance.

Story 2008, Erin Digitale. For reproduction requests, contact the Science Communication Program office for author's email address.



Erin Digitale
B.Sc. (biochemistry) University of British Columbia
Ph.D. (nutrition) University of California, Davis
Health Research Communications Award, Canadian Institutes of Health Research
Internship: Stanford Medical School news office

In grade three, I wrote a story about a girl hunting dinosaur fossils. On the first day, she found a tiny fossil. On the second day, she found a fossilized skull. And on the third day, she found an entire Stegosaurus skeleton!

My story, criticized by an 8-year-old classmate as unrealistic, foreshadowed why Im pursuing a career in writing instead of research. I dont want to spend years digging in the same pile of dirt, as a good scientist must. I want the chase, the climax, and the dnouementnow. Ive earned two postsecondary degrees via esoterica such as watching videos of wiggly enzymes, making rat-liver milkshakes, and lecturing undergraduates on the capabilities of the human stomach. Now Im eager to tell the stories of science beyond the academy walls.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Kacie Carter
B.A. (environmental studies and legal studies)
University of California, Santa Cruz

My childhood on the California coast filled me with a love of exploring the natural world through field science and artistic expression. I found the perfect outlet for my lifelong passions in the Science Illustration Program. Nothing is more beautiful or interesting than nature, and the attempt to capture a tiny fraction of that beauty brings me closer to understanding and being a part of that whole. Science illustration is a synthesis of my passions; to me it is an elegant combination of science, art, and environmental education. In the brush stokes of watercolor, or in a simple pen line, you can communicate the detail, form and wonder of the natural world. It becomes immediately accessible, and provides a gateway for understanding.