There is an accelerating flood of niche social news sites; it seems more likely to me that in a few years there will be a huge number of these kind of sites (and many platforms to let you make your own), which in aggregate will probably follow a power-law distribution of user populations.
Recommendations is a good idea (never figured out why it doesn't work at reddit), but only for the 'fat head' of the curve, for the long tail the problem almost takes care of itself. Also, better categorisation would be useful, and would help with recommendations too.
Sure, I was just saying that the lowest-common-denominator problem will mainly affect bigger sites, and more nichey stuff will be better served by topic-specific sites.
Interesting. You could almost s/reddit/digg/g and get the same article.
IMO, part of the downside of the whole "Web 2.0" social-whatever type sites these days is that they're so easy to create. Digg. Reddit. Mixx. Etcc. Why fix the broken one, when we can build a new one over a weekend (plus pizza and beer).
As the author outlined, they go through a standard-ish 4-step program as a sort of half-life dying off (at least as far as the hard-core geek users are concerned).
I postulate that you can't fix Reddit or Digg, and Mixx will be broken soon too. These sites start out by attracting the leading-edge hacker techie types. The ones who understand about a product being in beta (and feel a sense of elitism for knowing about it early on), they will report bugs to you (often with a 9 paragraph writeup of how to reproduce the bug(s)). They will submit good content, and comment on articles, and vote, etc. But they won't click on ads, or do other things that inch the site toward profitability.
So, there is no choice but to cast a wider net, gather in a lower-level audience. The folks that haven't yet seen a cat with a lime peel on its head (at least we're not getting pancake bunny anymore) flock to the site. With the masses comes the material that the tech crowd doesn't really care for.
news.YC might be the exception, if only because it's more hyper-topical, and it doesn't seem to be concerned with making money (I think news.YC is more of a flytrap for potential investees).
So, enjoy Digg and Reddit while they're young, hold on as long as you can through the growth phase, and then bid them farewell.
Maybe they will one day manage to recover. Maybe we'll read about their recovery here. Or on Slashdot...
Not to mention, unlike Reddit, they weren't neutral. They would try to squelch certain stories and ban users. They shut down all submissions after their attempted squelching of the HD-DVD hex key didn't work. They banned unpopular viewpoints. Censorship was de rigueur.
I think it was supposed to cluster users who often voted the same way, and then infer that if a user voted up an article, most people in the same cluster as that user would also like the article.
I wonder why it didn't work. They certainly have enough of a dataset to play with by now. I wonder if it's computationally too intensive to continually update each user's results, or too difficult to come up with a good algorithm.
Does this "recommended page" problem bear any resemblance to the Netflix prize problem?
Creating a good recommended page would certainly lure me back to reddit.
I don't think the founders have much of an incentive to change reddit a whole lot. Also, the existing algorithm is definitely computationally intensive: it takes 12 seconds to generate a new recommended page.
This post makes the mistake of assuming that reddit and similar sites actually implement "wisdom of crowds". They don't. They don't because you get to see the 'most popular' answers and choose from those to vote on, if you really wanted WoC to work you'd need to present people random stories to vote on. (This is part of what I attempted to do in my <a href="http://www.jgc.org/blog/2007/09/wildfire-has-launched.html">Wildfire</a> Facebook application).
He also proposes Bayesian filtering as a suitable technique for presenting the recommended page. Despite that fact that I'm a Bayesian-head (see <a href="http://getpopfile.org/">POPFile</a> for example) I don't think that's a total solution. You can do ML for news (it's an old idea, see the 20 newsgroups test set for example) and I have a private 'reddit' which I use as a feed reader that does ML.
But I think if you want to solve the 'reddit problem' you actually need to give people the chance to control the 'crowd' that they are getting recommendations from. I'd like to see (actually, I am building) an application that's reddit-like, but uses a combination of ML and social aspects to get a better view of interesting stories.
I have an idea for the Reddit people. The old people was the one you got when the community at Reddit was the old community, but the community now is a lot bigger, as there is a lot of people that there wasn't here then, and there is a lot of people that is gone too. So, imagine you want the old users to get the classic Reddit experience again, how would you do it? Just identify who the classic Reddit users were (that's easy, just check the time at which they did sign up) and let them access to a Classic Interface, where karma counts as if only they were part of the Reddit community, and nobody else.
That would make the new reddit behave as the old reddit.
I don't think so. The site has grown past a critical point and caused the quality of submissions to submarine below anything the "old" crowd has any interest in. To paraphrase a comment I saw on reddit: it has degraded to the lowest common denominator -- politics.
No. I'm sure plenty of the old users are still there. The problem is that there are too many new users polluting the community with less than worthless comments and submissions that are completely off topic from what was once the norm and driving force behind the community.
Reddit was once a great place, but its time came and went. Now there's n.yc, and whenever this site starts to drift off on a tangent, something else will crop up to keep us entertained. I think it's just the cycle of life for social bookmarking sites.
I have been working on an algorithm to solve the problem, but I'm not yet ready to release it out in the open. :-)
Baysian filters requires quite a lot of training to start working effectively. People don't want to rate hundreds of articles before they start getting good content.
Recommendations is a good idea (never figured out why it doesn't work at reddit), but only for the 'fat head' of the curve, for the long tail the problem almost takes care of itself. Also, better categorisation would be useful, and would help with recommendations too.