20bits: Implementing A Suggest-A-Friend Feature | 20bits
Matt Rubens
· 1 year ago
Here's the problem: this algorithm only predicts if someone is in your social graph -- it doesn't predict whether you like them or not. In fact, I would argue that past a certain point, having a lot of mutual friends negatively correlates with the probability that I want to friend someone.
I've joked that Facebook should rename their feature "Ex-girlfriends and people you hate"...
Jesse
· 1 year ago
This is definitely true, although sometime in the last few weeks Facebook added a feedback feature so you can banish people from your suggest-a-friend list.
I'd say I banish about 90% of the people who appear.
SG
· 1 year ago
Networks like Facebook have more data than just the 'friend of' edges. They could use the interests, favorite movie/books, groups etc to better predict if two people not yet directly friends have enough in common to 'like' each other.
Matt Rubens
· 1 year ago
Yeah, me too. And there's something cathartic about the banishing...
It's interesting to think about how you could improve the algorithms to take like/dislike into account though. For example, if both the other person and I have been on the site for a while and have large/stable friend graphs, we probably don't like each other. But if either of us are new to the site, it's probably worth making the suggestion.
Jesse
· 1 year ago
SG,
I don't see how interests listed on Facebook have any bearing with whether or not I'm friends with them.
Besides, improvements in the algorithm would come from less locality, not more. As in the HRG approach you'd want to start learning about relationships between cohorts and components, not between individual people.
Facebook's feature is "good enough," anyhow. The main goal, IMO, isn't to improve their copy of the social graph. Rather, it's to (1) improve the experience for people just joining the site as they add their first seed friends and (2) reengage inactive users by emailing people when potential friends join the site.
All Comments
I've joked that Facebook should rename their feature "Ex-girlfriends and people you hate"...
I'd say I banish about 90% of the people who appear.
It's interesting to think about how you could improve the algorithms to take like/dislike into account though. For example, if both the other person and I have been on the site for a while and have large/stable friend graphs, we probably don't like each other. But if either of us are new to the site, it's probably worth making the suggestion.
I don't see how interests listed on Facebook have any bearing with whether or not I'm friends with them.
Besides, improvements in the algorithm would come from less locality, not more. As in the HRG approach you'd want to start learning about relationships between cohorts and components, not between individual people.
Facebook's feature is "good enough," anyhow. The main goal, IMO, isn't to improve their copy of the social graph. Rather, it's to (1) improve the experience for people just joining the site as they add their first seed friends and (2) reengage inactive users by emailing people when potential friends join the site.
Also, remember, more data usually beats better algorithms.