3 ms·
Lots of good, interesting ideas in that post. A few comments: In general, what data you capture, always depends on what you want to do with the data. Any dat
by feral 15y ago
Lots of good, interesting ideas in that post.
A few comments:
In general, what data you capture, always depends on what you want to do with the data.
Any data model is an abstraction, which discards a lot of information. This simplification is what makes the data model useful.
Of course, storing social relations as an undirected, graph with a single type of edge, as is done in Facebook, is a gross simplification of real social structure. Everyone knows that; Facebook know that; but the fact is that you can build a lot of useful stuff, with even such a simple model. Is it perfect? No - I'm sure we will do better in future - but you've got to admit its amazingly successful, for such a simple representation.
So, we aren't going to be able to define a format for the social graph that is so rich, that it will capture all possible uses with it - but I don't see why we couldn't define a format that captures enough detail to do many a great many of the things we might want to.
Next point: graphs are amazingly flexible data structures; if you have a multigraph or hypergraph representation, or a set of graphs, you can represent an awful lot of rich information. Don't knock graphs too much. You could build an awful lot of cool communication functionality, if you had really any rich social graph representation, not necessarily the 'perfect' one.
The author also writes: "In other domains, a big graph would be good for recommendations, but friendship is not transitive. There's just no way to tell if you'll get along with someone in my social circle, no matter how many friends we have in common." Of course, friendship isn't strictly transitive; otherwise the giant connected component of a social graph, would be one big clique; but friendship is highly transitive. That's why recommendation works so well in Facebook. If A is friends with B, and A is friends with C, then there is a vastly higher chance that B is friends with C, than with some other random person. Sociologists have called this 'triadic closure' and studied it for years: http://en.wikipedia.org/wiki/Triadic_closure http://en.wikipedia.org/wiki/Triadic_closure
So, while the recommendations won't be perfect, as things like friendship aren't perfectly transitive, they'll potentially be very good. But what recommendation engine is perfect, anyway?
Finally, the author refers to the problem of finding an adequate model as 'AI-hard' - well, as I've said, any model - even the most perfect one (unless we go and make a copy of the entire state of the world) - is going to sacrifice something. So I see why coming up with a perfect model is impossible. But as to coming up with a model that is adequate for a great many of the possible things we might want to build; well, I'd say 'AI-hard' is an exaggeration.
- andrewflnr 15y agoFacebook has a heck of a lot more info on relationships than an undirected graph with a single edge type. In the front end alone, they now have a directed "follows" relationship. In their backend, they're keeping track of how much you pay attention to people, comment on their posts, click their links, etcetera, and using it, among other things, to decide what shows up in your news feed and what is silently dropped. Someone posted a link here not too long ago describing how to get the weights Facebook ascribes to your social connections, based on the above sort of metric.
- feral 15y agoYes, of course they do; they have vast amounts of data; click streams, likes, comment text, etc. But I guess I'm saying that the core metaphor, with which Facebook did much of its early growth, was that two people are 'friends', or are they are not; and that you see the activity of your friends, and sometimes see activity, and get recommendations, from the friends of your friends. They built many iterations of the product, and got a very long way, by exposing a very simple social graph metaphor to their users. Even though that model drastically simplifies the world, it has proven to have a lot of utility for users. Maybe now, because the users are consequently more educated about social networks, a more sophisticated metaphor would go further.
- rboyce 15y agoThanks for mentioning triadic closure. It's a great example of how useful information can be inferred/suggested based on an incomplete or imprecise model. The sacrifice of simplifying the social graph is what makes the task of constructing this graph online approachable.
- ableal 15y agoI hope Maciej is not going soft on the trademark "brevity is for the weak" style, which I've been appreciating since idlewords.com days. I believe he spared the spelling out of the LOAF acronym, and scanning the very interesting http://loaf.cantbedone.org/about.htm http://loaf.cantbedone.org/about.htm did not clear that up. "List of a friend"?