Hacker News Reader

Comment by TheOtherHobbes | Hacker News Reader

The problem with habit reinforcement is that you've simply created a feedback loop.

"You like [music type]? Here are more bands/artists of [music type]."

It looks like an innocent service. But it's devastating to real exploration, because it makes it much less likely you'll ever discover Band/Artist Z whom you'd never normally listen to but love anyway.

This is why thoughtless customer profiling is a dumb idea, and certainly not the sure fire insta-profit marketing panacea it's sometimes supposed to be.

It has some uses, but you're reducing customers to stimulus/response robots with a limited behavioural repertoire, and that's an excellent way to miss a lot of opportunities.

View on HN

> The problem with habit reinforcement is that you've simply created a feedback loop.

> "You like [music type]? Here are more bands/artists of [music type]."

So far I've found the opposite. Pandora has exposed me to music that I never would have sought out on my own. In fact I can think of several artists that I would have judged by their cover, so to speak, and never given a chance even if I had stumbled on them on my own, instead of coming to awareness mid-song that "I don't know what this is, but I kind of like it."

My wife teases me about this a little bit when she finds me listening to music that she says isn't 'me' or doesn't seem like something I'd like. She's right, but I guess Pandora knows my tastes better than she does (and better than I do, to be fair.)

Contrast that with Spotify, which when given a song or artist that I like goes out of its way to recommend music that is superficially similar but that I hate. Or how it can recommend 80 songs to me a week based on things I have listened to and end up recommending me nothing that I like.

Huh. I've actually found the "Discover Weekly" feature to be remarkably good at picking music that I like. I'd say I affirmatively like about half to three-quarters of the songs per week, with only a couple that I find myself skipping.

I wonder if it's better at certain genres than others.

> I wonder if it's better at certain genres than others.

The Discover Weekly algorithm is, as I understand it, based on what songs have been added to other user playlists. So if you're primarily listening to a genre that has a lot of intense people making carefully curated playlists, you're gonna get a better Discovery Weekly.

And more generally, I would say it just depends on the quality of the AI. There is no fundamental reason why such a system would have to recommend only similar music. It could try contrasting music, and build a picture of what aspects of the music you like. It could draw conclusions to recommend seemingly different genres which nonetheless share some commonality that might make them appeal to you. It could also occasionally try out completely different things, to broaden your horizons (and give you more of that if you react favourably). And that's just scratching the surface. You could do all that without AI. An AI could potentially do all that, plus other interesting things that we wouldn't even think to try.

The algorithms used to make musical recommendations do statistical analysis on numerical ratings given by different people to different songs. They typically know nothing about music, as musical knowledge is understood not to improve their results. I wouldn't call them AI - they're neither knowledge-based systems nor neural networks.

I don't think that that's true for Pandora: https://en.m.wikipedia.org/wiki/Music_Genome_Project

> I wouldn't call them AI - they're neither knowledge-based systems nor neural networks.

Out of curiosity, why are you drawing the AI line at neural networks? Neural networks are just function approximators; there's nothing magical about them except they work really well on problems we're interested in.

Would you really not consider an SVM or a Naive Bayes classifier "AI" since they don't use a very rough, very loose analogy to the brain?

I discussed this problem decades ago when I developed MORSE (http://web.onetel.com/~hibou/morse/MORSE.html).

People like certain things for no obvious reason, and are unlikely to change. It does them no favours to recommend them music they don't like, or anything else which depends on taste. But if there's music similar to the music they're known to like, which they haven't heard - perhaps because the band is obscure, from a different decade or part of the world, or new, for example - they might want to hear it. Also, there might be one or two songs they'd like by artists they don't normally listen to. Collaborative filtering is good for this.

Your point is valid about ideas, news, and a number of other things which have nothing to do with taste. People would benefit from sometimes being exposed to viewpoints different from their own.

If you however do succeed in reducing humans to stimulus/response bots, it will likely be good business. Which is the scariest aspect, cause that drives lots of decisionmaking today...

Hence random mutations in genetic algorithms. And other similar functionalities

Reply on news.ycombinator.com