Comment by bravura | Hacker News Reader

Comment by bravura | Hacker News Reader

I've thought about this problem a lot, and I think the problem is recommendation systems (e.g. for feeds) are not optimizing for awesome. They optimize for having an endless amount of decent.

Bob Carpenter of Alias-I wrote a blog post about a particular NLP task he was working on, and he decided that the correct evaluation measure was to insist on 95% precision (must be awesome), and only then optimize recall (amount of awesome discovered).

In the real world, I have a few friends that don't make recommendations really often. But when they do, I jump. Their recommendations are always solid gold.

The problem with recommendation on the web is that there is no built in notion of economy. Economy in the sense of "frugality in expenditure or consumption". There is an inherent economy on the amount of time I have to consume content. So why shouldn't there be an economy on the amount of things I can endorse? Perhaps that would guide content sharers at large to only share awesome. (And I get more endorsements based upon how much people value my previous endorsements.)

I'm toying with calling this idea "the economy of awesome" or "the economy of cool".

> In the real world, I have a few friends that don't make recommendations really often. But when they do, I jump. Their recommendations are always solid gold.

Can't we solve this by allowing users to internally rate their friends? Just like you, I only care about the recommendation of some of my friends. So if can somehow rate my friend's recommendation from 1 to 5, then the algorithm could just to only recommend me stuff from the friends I value the most.

You can even use machine learning to do this automatically. Instead of bothering users to rate friends, if I often like the recommendations you give to me, I should get more recommendations from you. If I often ignore your recommendations, I should get less of it.

G+ circles could even be used for further context. I might like your recommendation for food, but not for tech. So the engine could could separate each content with a context (this is a tech blog, or that is a food blog) and relate that to my circles. So I would get tech recommendation from people on my tech circle, but not from people on my food circle.

I think the big picture problem here, is that social networks usually put all 'friends' in the same bucket of equal value. When in the real world we have different values for different friends, and for different contexts. So a solution would be to simply have additional variables for friend rating, then use machine learning to try to automatically match them to each context.

This is a partial solution, but one that is already implemented.

The problem is that people have different semantics for clicking "Like" on something, or giving it five stars. This becomes problematic when giving something five stars means that it's at least "pretty good to you". In that case, you lose the dynamic range of expressing "this is AWESOME. simply AWESOME."

So a machine learning algorithm cannot distinguish between good and AWESOME if the user is lumping them into one category.

Which is why I propose actually creating scarcity around the amount in which you can promote or approve or content, perhaps backed by a currency of cool. I can only click like five times a day, for example. That would limit people's ability to blithely approve of everything, and their ability to collapse the distinction between good and AWESOME.

This approach is much stricter than the optional "gentleman's agreement" proposed in the article, but might also create more desirable properties, and a fun incentive around only sharing AWESOME.

Can't we solve this by allowing users to internally rate their friends?

I think Facebook already does that. Notice they put lists in your profile now called "Friends", "Close Friends", "Acquaintances", etc.

The one thing that annoys me about this notion is that those relationships change IRL. What if I suddenly want to have a party and invite only my "Close Friends", which then ends up including some guy I very much hate by now?

One of the biggest problems with social network evaluation schemes is that they always work on the assumption that relationships between people are static. They are, however, dynamic. And the rate of dynamics changes from person to person and changes over time as well. It's not as simple as saying "this guy makes the best content ever". Maybe tomorrow his dog will die and he'll start posting emo poems.

Wish I could upvote you twice for rephrasing what has been ruminating in my mind for such a long time. The problem is an economical assymetry: it 'costs' far less to produce content than to consume content so there is no incentive to be selective. What if we somehow could artificially constrain the supply-side?

This is exactly what I'm proposing. The idea revolves around creating a scarcity. I'm inclined to call it an artificial scarcity, but in some respects the scarcity is intrinsic in people's finite time.

Simply put, you start with a certain number of "cool dollars". You spend them by promoting something. I'm still working out the mechanics so that the market has the desired properties. Maybe you get one dollar a day. Maybe you spend money to agree with someone's endorsement.

Minor quip: assymetry is the science of measuring mules.

ranking social content created and shared and filtering out the noise is I think the biggest problem with current gen of social networks and though fb's edgerank is better than anything else out there I still find myself on g+ to avoid the overload. but the biggest offender is twitter using only time as the decision making point to update me about my interest graph is not very effective (and its slow loading of old tweets doesnt help either). [though if its only intended as a real time feed I may be using it wrong]

The biggest problem is that Google only had to index HTML; there isn't even an all-encompassing standard for social networks, and if there were, it would quell the current rapid development we're seeing. You don't want to freeze something in place before it reaches maturity, and social networks definitely haven't.

However, perhaps instead of exposing the data, you could expose the search better. Then, specific standard APIs could be defined concerned with what you want to search for: for example you want to find all the neighboring nodes of a specific node, or during christmas find all people who are friends of sport enthusiasts. This would, however, take the control away from the likes of Google and put it in the hands of the content holders, such as Twitter and Facebook.

Reply on news.ycombinator.com