Hacker News Reader

Comment by nkwiatek | Hacker News Reader

"People who live in Palo Alto who like Sushi who aren't my Grandma"

It's the future

ITYM "People who live in Palo Alto who like Sushi who aren't my Grandma or my one friend bob who has horrible taste or oh crap that one guy fred who i accidentally friended or ..."

I don't know anyone who has a very clean social graph that they can formulate sane preference queries about.

Indeed. Which is exactly why you're better served using a dataset that includes people you don't know (and ideally who aren't like you or else everyone will segment off into the same couple spots like middle school). The wisdom of crowds doesn't mean the wisdom of your rolodex.

I did at one point organize my FB friends into groups like Family, Close Friends, Acquaintances, Distant Friends, People I Know From That One Club, etc. If I were going to use this, I'd probably just filter it to use the Close Friends group. Or if I know I have similar taste to one or two friends, I'd just limit it to them. I dunno, I'm not sure how I feel about this yet, but I can certainly see some ways I could use it.

How much information on sushi places in Palo Alto is your grandma pumping into the system to dominate that sector? Is she checking into multiple sushi places dozens of times a day? It's like supplementing every google query with -site:4chan.org on the off-chance a search for a camera lens spec or node.js tutorial would lead to 4chan.

Replace "grandma" with "person whose taste in sushi is very different" if it makes you feel better.

Mmmkay, so out of 3 ways to implement ranking:

1) Rank by popularity among all friends.

2) Rank by global number of check-ins.

3) Rank by a one-off friend's preferences.

You think (3) is the way something called GraphSearch works?

This is a great strawman that you've knocked down.

I think it's closer to #1 with "popularity" replaced by "preferences". But that has the problem that i don't care what all of my friends think about most things, or even most of them.

Facebook can solve this algorithmically without you needing to have a "clean social graph" or write complex queries to exclude your grandmother.

Proof? Most of these algorithms build models that do affinity weighting, and then try to guess at how the same some other person is as me (which is hard across such a variety of attributes).

There is no magic, so please explain the math/algorithms you think would work here.

(Sorry, it's just a lot of people wave the magic "algorithm" flag when faced with hard problems)

Well, Facebook can tell how often you interact with certain people. They can determine certain things about the content of those interactions. They can detect how similar your friend graph is to another person's. They know that your grandma is your grandma (if you have told them, of course), and they know, for example, that most people don't want to hang out with their grandma.

I haven't given you pseudocode for an actual algorithm, but I can imagine that Facebook could combine all of these metrics into an index that can tell them roughly how likely you are to want to hang out with certain people.

without you telling explicitly who you want food recommendations from or tech product recommendations from its an incredibly difficult problem. Just because you talk to someone everyday doesn't mean you want their opinion on food or tv shows.

Their "Close friends" algorithm does a pretty good job already. And even if it doesn't I'm sure most people have already removed / added friends as appropriate from that circle.

Facebook also has a bunch of data on how often you've interacted with various friends there .. so that could be one data point. (How often you're tagged in a check-in / photo, how often you've liked a post made by someone else etc.)

Your Your mom / dad / school friends generally may not be a good data point, while your college friends might be.

It'll take a lot of beta testing and tweaking, but I think it can be done.

You're not querying the data to get one answer. You're harnessing existing data to narrow down hundreds of friends into a handful.

Reply on news.ycombinator.com