Monday, June 4, 2007

Context And The Crowd Trump Behavior

by Steve Smith, Friday, June 1, 2007
MOST BEHAVIORAL TRACKING WORKS ON the simple assumption that past behavior is a solid predictor of future need. Not so in all applications, says Baynote CEO Jack Jia. The founding CTO of content management giant Interwoven took a crash course in social science, which he claims tells us that group behavior in a given context is a better indicator of any given user's future actions than piles of personal behavioral history. He tells us how the wisdom of the crowds in context informs Baynote's approach to content and product recommendations.

Behavioral Insider: You were at content management supplier at Interwoven for eight years. How did that experience evolve into Baynote?

Jia: The last years of my Interwoven experience really helped me start thinking about the bigger problems people were facing. Customers were saying, the more content I am publishing, the less valuable it becomes. People aren't finding it. I created all this stuff but they aren't really getting value. This became the genesis of this new company that helps people consume this content more effectively. It follows how information and knowledge is accumulated and how it is passed down generation to generation.

Behavioral Insider: Explain that connection to human behavior patterns.

Jia: I actually spent almost six months studying and working with some of the social scientists at Stanford, really thinking through to the right solution -- not from a tech or computer science perspective, but more from psychology, sociology and organizational behavior.

We produce content recommendations. Where other people find value with similar things on your site, we start to accumulate that knowledge by observing their behaviors. We can deduce what is useful and what is not. and feed that back to the users with similar needs. The bottom line is the conversion rate improvement for any business site. Our e-commerce customers have seen somewhere in the area of 20%-plus revenue lift. The support side is cost reduction. Fewer people call if they find the answers online. We just did Turbo Tax's support site and they saw a huge reduction of the support costs by having the community being the backbone of providing what answers to provide to which questions.

Behavioral Insider: Explain that connection between the social science you learned at Stanford and the company's behavioral model.

Jia: We found what we call the wisdom of the invisible crowd. This is a big insight that the psychologists and the social scientists have known for years. There is a controversy between behavioral targeting and contextual targeting. A lot of computer scientists believe you can predict people based on their past behaviors, that past needs can actually dictate what I want in the future. And that's really the basis of personalization and the more recent BT.

When we went to Stanford and really studied this, we found [past behavior] is a very poor predictor, because humans have way too many profiles. I am a father, a son, a brother, I like travel, I like a lot of things. You can track all my past behaviors all you want, but in any given moment when I go onto a site it is very hard for you to predict what I want. We believe that particular approach has a low accuracy in terms of predicting human behavior.

However, there is opposite research that is well known among scientists that asks whether people are unique. We are raised with the notion that I am unique and don't have needs quite like other people. But the scientists proved we are not unique at all. We are pack animals. Pretty much 95% of people will need the same thing. We only need to find out under what context what things are useful, and then present that product or content given the context. Then the prediction is very accurate.

Behavioral Insider: Describe how that insight translates into a recommendation method.

Jia: For example, although we know each individual and track them, we don't really care what they did yesterday. We can build profiles, but generally we don't use that primary guidance. Instead, say at our customer Cisco's site, someone goes to the router portion of the site -- then you will more than likely behave like other people who have interest in routers. There is a router peer group vs. the switches Cisco also sells, or the investor visitors vs. the job seeker. They can be the same people with different profiles, but once you enter that router world you behave much like the other people in routers.

We don't really care about your behaviors. And given our attention is so short, we can only recommend to you three or four documents before you get impatient. And then which ones you choose is very important. In aggregate, you are watching everyone's behavior. But it's not necessary to track behaviors day after day.

Behavioral Insider: But what data points are you collecting to understand what people will do in a given context?

Jia: That's one of the unique things. Almost all of the vendors out there except Amazon is collecting clicks. Through our studies at Stanford we found the click doesn't tell you anything. The click is the function of the area of the site you are on. If you put a link on the front page, you will get more clicks. It is a self-fulfilling prophecy. If you link less-used content on the front page more people will click on it, and then people back out of it.

So click doesn't tell whether a piece of content is useful or not. It is the other things we are tracking, about 20 behavioral characteristic that in aggregate can tell us, for instance, how much time you spent on a piece of content relative to the size of the content. It really depends on what other like-minded peers also spent time with. Whether you come back to the same content. A repeat visit to the three or four products you like, we call virtual bookmarking. So, basically how you navigate now is telling.

Behavioral Insider: What kind of sample base is needed to make predictions?

Jia: One person can be noisy. Two people still noisy. Three to 10 with similar needs, like-minded peers, will start to cancel the noise out. By observing a set of heuristics, that forms a unique fingerprint of someone who likes or dislikes the content. Then we compare that fingerprint with like-minded peers, 10 or 100 other people. The noise will cancel each other out in 100, and what remains is a consistent pure signal. That is where we know that this content is truly important to that group of people.

No comments: