Great Recommendations Are Half Placebo • Hacker News Board

Re: Great Recommendations Are Half Placebo

#11

Post by shokwave » Tue, Mar 05, 2013, 7:17 PM UTC

The author writes as if, when a 'real person' makes a recommendation, they aren't just making a prediction from data. The difference is that we can see how the algorithm works, whereas our picture of what the human brain does when it thinks about what someone might like is very murky.

And even if trusting a recommendation causes a placebo effect that increases the perceived quality (as others have pointed out, it seems plausible that a hyped recommendation may cause disappointment instead), why on Earth would we trust a process we don't understand, but mistrust a process we do?

Re: Great Recommendations Are Half Placebo

#12

Post by 1rs » Tue, Mar 05, 2013, 7:54 PM UTC

I think the 'black swan' is important in recommendations as well. There's stuff that will blow your mind because it's something you think about and is somehow connected to something you've done. It might not be relevant at all to most people who read the same stuff as you do. So it's very unlikely it will be recommended to you. Recommendations increase 'groupthink' and you find interesting stuff but not a lot that's outside of your comfort zone. Humans are still better at connecting dots and giving recommendations than machines. I guess it's difficult to teach a machine to make great surprises?

Re: Great Recommendations Are Half Placebo

#13

Post by biscarch » Tue, Mar 05, 2013, 8:06 PM UTC

nathan_long wrote: post #6

A great recommendation engine should incorporate explicit feedback from users. For example, Netflix knows what I've watched and how I've rated things. But it doesn't let me tell it things like "I categorically dislike horror movies", so it sometimes recommends those. It's trying to guess something that I'd rather just tell it.

Wouldn't the problem then be what you consider horror vs what other people consider horror?

What if it's labeled as horror and comedy because it's a parody?

Re: Great Recommendations Are Half Placebo

#14

Post by nathan_long » Tue, Mar 05, 2013, 10:33 PM UTC

eli wrote: post #10

nathan_long wrote: post #6

A great recommendation engine should incorporate explicit feedback from users. For example, Netflix knows what I've watched and how I've rated things. But it doesn't let me tell it things like "I categorically dislike horror movies", so it sometimes recommends those. It's trying to guess something that I'd rather just tell it.

Isn't that what this does? http://movies.netflix.com/TastePreferences

Oh, I never saw that! Thanks!

Re: Great Recommendations Are Half Placebo

#15

Post by calhoun137 » Wed, Mar 06, 2013, 1:40 AM UTC

This article made me laugh, let's review.

Computer predictions seems to work, but only get you half way there. I don't trust them because they are from a computer, but if I wrote the program myself I would trust it 100 times less. As with baseball, guessing the home team will win works at least half the time. Our recommendation engine works, and people read articles we recommend 4 times longer.

Conclusion: computer based recommendation engines are 50% placebo.