Hacker News Reader

Comment by crasshopper | Hacker News Reader

Maybe recommendation engines fail just because the math is crappy. Current scholarship in the field is pretty shallow and/or hackish.

View on HN

The best algorithms in the world won't make up for a lack of data. Few companies are in a position to make good recommendation engines.

I also don't think math is to be blamed -- human psychology plays an important role -- i.e. I may be interested in software-dev stuff or in startups stuff, but not every hour of every day. Basically I'm inclined to read such stuff only in the morning, while drinking my coffee (the perfect time of day for big plans).

But in the afternoon I prefer reading articles about my OTHER (lighter) interests / hobbies, like photography. And I also prefer getting stuff solved, like finding an electrician to fix the poor wiring in my house or searching for gifts for my wife/child.

For any person on this earth, immediate interest in something is relative, based on time of day, time of year, mood, current needs, current problems, current hobbies, etc...

A good recommendation engine is practically unfeasible -- because if it cannot anticipate your future desires, both short-term and long-term, then it's useless; because people "search" for stuff as soon as they realize they have a new need/desire, and then it is too late to make a recommendation (at this time, the user's history is only useful to "understand" / parse the search query).

Recommendations need to happen before the user searches for stuff.

For example -- I searched for books on drawing / colors on Amazon, and bought 2. Amazon should have anticipated that I may like getting into photography, as they do know a lot about me, but it didn't, even though Amazon's recommendations are some of the best. Now that I already bought stuff, I don't care about special offers on the same kind of stuff that I already bought.

Also, people with "hackish" attitudes have done pretty amazing stuff btw, contributions to AI included -- software isn't just math, it's also biology / chemistry / physics / psychology / philosophy, amongst others.

There is no reason why time of the day, temperature outside, whether Obama made a joke or not, etc can't all be features for a classifier to learn.

I love the idea of an algorithm anticipating my next move and I'd argue that for the vast majority of people you can find repeating patterns that are almost always relevant. These patterns will not be simple or obvious and as you currently pointed out, the lack of data is what is holding most innovation behind. I feel that companies that own the data like Amazon will have a lot to gain if they do a Netflix-style competition to improve their recommendations engine.

assiotis, have you tried giving a bazillion data points to an SVM? One needs a sensible kernel as well.

I agree completely if we're talking about a recommendation engine that's supposed to always be able to find something of immediate interest for you. However, that problem is at least as hard as building a general AI, since not even my friends can always think of something I would find immediately interesting.

If you relax the scope of what you want a recommendation engine to do, they can do an awful lot. Directed Edge is absolutely killing it here -- http://directededge.com/ . By providing a white-label API so that developers can specify the data on which to base recommendations and by using some superior algorithms, they're able to come up with impressive results for recommendations like "People who might like to buy this might also like to buy this", or "These users might be your friends", or "You should read this other article."

bad_user, how is what you mentioned beyond mathematical modeling?

Something else you may have neglected in this response: the software doesn't have to just make a recommendation. It can present you an array of options which you filter. A dynamical interface.

Most of them fail because they don't have enough data, or they're using the wrong kind of data.

The main assumption with social recommendations is that I give a damn about everything my friends like. With every one of my friends, we have an overlapping interest in at least one category, but but there are about two people in my social network that have an overlapping interest with more than half of my interests.

With current engines, they say "Oh, Alice liked the new Katy Perry album, you should check it out", without any understanding that I'm friends with Alice because of she can hold an interesting conversation, rather than a shared music taste.

The proper way to do recommendations is to build up a profile of the user, and then match recommendations from people similar to the user. This is exactly what Hunch is doing. I hope they're successful, they're the only ones close to a solution.

(Disclaimer, I'm not affiliated with Hunch, but I'm founding a startup using very similar math in a different field).

I agree that data is the main issue here. Most companies using recommendation algorithms are over-relying on math to do the magic and make up for the fact that they are blind to what they are recommending. In order to recommend you photography related items based on your past interest in painting, the service would have to understand an awful lot about the two realms and what they have in common (art, visual, etc). Algorithms alone can't figure this out. No, I'm not blaming math, math can do amazing things to help us deal with situations where we have no data. But meaningful, accurate recommendations require meaningful, accurate information about the items being recommended. IA techniques can help automate the cataloguing of this information, but 'blind' algorithms alone will never make high quality recommendations - the kind that make you say 'Yes! Exactly!'

I aslo agree that social recommendations have to be able to pinpoint which friends have simialr taste, otherwise the result is just noise and junk.

Nikki www.jinni.com

Nikki, I would argue that it's not a case of "maths in general can't do this" but just that you may be using the wrong maths. You can't "just apply SVD" and expect that to work super well.

I feel like a good recommendation engine would be based on browser history. Google would be a in a prime position for this.

...what is a search engine if not a recommender of content?

This would be better suited as a browser extension.

Feature, not product. No-one cares about recommendations by themselves, they just want to buy or look at stuff.

(DirectedEdge's product is, effectively, developer tools, so people who want the feature don't have to build their own. Hunch is heading that way too, looks like.)

adw, no-one wants iron ingots either but that constitutes a finished product.

Reply on news.ycombinator.com