CHI Paper: Predicting Tie Strength with Social Media
[Sketchy guy from high school] added you as a friend on Facebook. We need to confirm that you know [sketchy guy from high school] in order for you to be friends on Facebook. Confirm, or ignore?
Eric Gilbert and Karrie Karahalios presented the largest regression analysis at CHI, and it was focused on just this situation. Can we build a model to predict how close two Facebook friends actually are, using only data that Facebook has access to?
So, can we separate sketchy high school classmate from your BFF? Yes. Eric used a linear regression approach common in social sciences; he and I think that you could do even better using machine learning tools available today. The upshot of a linear regression (as opposed to an SVM kernel) is that you can inspect what impact each variable is having on the relationship.
So what do you want to pay attention to if you’re Facebook or a Facebook application? The top five:
- Days since last communication: if you’ve recently written on their wall or sent a Facebook message, chances are you’re good pals.
- Days since first communication
- Intimacy and Structural interaction effect: This means, roughly, that a friend clique between you and the other person modulates how much it matters when you chat a bunch on Facebook.
- Wall words exchanged: “hey michael was great to see u u were so drunk last nite lol omg im still hungover.” Yes, the more of these rambling wall-to-wall posts you have, the closer you probably are.
- Mean strength of mutual friends: if your mutual friends are really close, then you probably are too.
This work won Best Paper at CHI. Congrats Eric and Karrie!