: 2008 August :

Yesterday alone, Facebook users issued 21 million friend requests. 17 million requests were accepted. So many new connections, and yet they’re all treated the same—what an oversimplification!

All Facebook links are created equal. But links can differ in strength—for example, a close friend versus a casual acquaintance. Links can be in different categories, like your boss versus a random hookup. And links can be asymmetric—Amy may think that Bob is a good friend, yet Bob may not trust Amy at all! The world is not a binary place.

Discovering Discovery: Don’t ask, Do tell

How can we use data to investigate these different properties of links? Today’s social networks do a lousy job of leveraging our existing data. Why do you need to manually confirm my friend request if we’re already calling, IM-ing, and emailing each other all the time? These data sources should be able to make a good guess about the strength and type of our relationship. Why not use existing data sources to propose better default responses?

If we give our networks a richer structure for our links and relationships, we will also be able to discover interesting facts about ourselves. Why is this important? By investigating implicit relations, we can gain insight into our relationships and how they work. For example, I might be surprised to find out that whenever I email my friend John, he always writes me back promptly whereas I always take 10 times longer to respond to him! Armed with this knowledge, I would ask my system to tell me to get my act together and crank out that response if I’m getting too delinquent.

Facebook 1.0 has helped us create an intimate network of our 17,000 friends. Will Facebook 2.0 help us manage them?

Mind the Explicit, Mine the Implicit

What else can data tell us about the quality of our relationships? One way to use data is to figure out differential interest in budding relationships. It’s easy to do this by looking at communications patterns in email, for example—does one person spend hours crafting that perfect email, only to get a reply that took only a few minutes to write? Or has he suddenly acquired a brand new set of favorite books, movies, and music that just happens to match his new love interest? People leave rich traces on the web—we can discover much more about them than the data they explicitly give.

This is only possible if we can look at the user’s history. After all, we can only make inferences about our behavior if we have a past to compare it against. But this introduces new questions: how much would you pay to know how long Monty spent writing you that email? How much would you pay to keep your data private?

Trust Networks

Social networks are also great for learning about trust. Let’s say that I’m thinking of entering in a business deal with you, but I don’t know you too well. Should I trust you?

There’s an easy way to use the power of networks to answer this question. Let’s just look at all of your other connections: do they trust you? We can give people reputation scores by allowing users to rate their interactions with friends. To make the system even more powerful, we could allow users to link their reputations. To illustrate: let’s say I trust my friend Mike so much that I am willing to attach a trust coefficient of 0.9. This implies that if Mike’s rating goes up by 1, I should get a rating boost of 0.9. Conversely, if someone has a bad experience with Mike and downgrades his rating by 1, my rating will also go down by 0.9. Through the power of the community, reputation ratings would spread quickly. (What trust coefficient would you attach to the author of this post?)

Reward Content Generation

One of the best ways to engage users is to get them to understand how every bit of data they contribute will end up benefiting them. In the example of trust networks, people can improve their own reputations by linking themselves with others. In my previous post on communication, I talked about a system where providing feedback on an email’s relevance would directly benefit you in the future. Online social networks need to reward people to provide explicit data, too.

The Facebook Feed was a brilliant idea for surfacing relevant content created by friends. Ideally, the Feed would create a positive feedback loop: good content provided by friends would get high ratings, which would motivate them to post even more good content. However, an early system of allowing users to rate the submissions of their friends was poorly designed—only 21% of users used the feature. On a rainy day, April 15, 2008, Facebook turned off the feedback system. What a step backward! I wish Facebook instead had created a better machine learning system to reward its users to generate and surface good content.

Social networks based on mutually confirmed binary relations was Day One in evolution of social networks. Introducing, richer semantics, more expressive structures including trust coefficients are the beginning of Day Two. What will the second week bring?

(0) Comments    Read More