3 ms·

>We've positioned machine learning as a replacement to curation, but what it really does is put forth the least objectionable content. I don't think the proble

>We've positioned machine learning as a replacement to curation, but what it really does is put forth the least objectionable content. I don't think the problem is with ML. You could easily train a net to make boulder selections with an appropriate data set. I believe what you are seeing is an accommodation for the average person. Most people aren't adventurous, and it makes business sense to get algos working for the middle majority than to cater to tail ends. It also, IMO, makes content at places like youtube less technical and more clickbaity. But again I believe this is somewhat intentional.

AlexandrB 8y ago
>> We've positioned machine learning as a replacement to curation, but what it really does is put forth the least objectionable content. > I don't think the problem is with ML. You could easily train a net to make boulder selections with an appropriate data set. I believe what you are seeing is an accommodation for the average person. Most people aren't adventurous, and it makes business sense to get algos working for the middle majority than to cater to tail ends. You literally just said the same thing the GP did while ostensibly disagreeing with him: ML curated content caters to the lowest common denominator - i.e. the content that most people would be okay with.
Declanomous 8y ago
> boulder selections Would that mean more rock and less bluegrass? (I'm kidding, of course) In all seriousness though, the problem probably is the data set to a certain extent. I am pretty adventurous, true, but at the same time I think the problem is that regardless of what data we collect, we only have the ability to determine if something is inoffensive. All you need to do is look at a site that aggregates reviews like Yelp or Amazon, and read some reviews. 95% of reviews are either 1 star or 5 stars, and it's obvious from reading the text that the rating is meaningless. "I ordered the wrong item by mistake, 1 star" "Food is overpriced, but the bathrooms are clean and wait staff is attractive, 5 stars." I find star ratings most useful because people who rate things 2-4 stars generally have the most nuanced and productive things to say about the product, and actually rate things what they think they should be rated, rather than using the rating system as a vote for "rate this higher/lower." I don't think it's possible to actually come up with good recommendations based on user-reported like/dislike rating. It's not wrong for a user to dislike a song because it reminds them of an ex, but using that as a basis for a recommendation to someone else is entirely useless. Systems like Rotten Tomatoes works really well in this regard, but almost has the opposite problem, which is that it tends to underrate movies with broad appeal, but that's generally not a problem since users will be exposed to those movies anyways.
tomatotomato37 8y ago
>Most people aren't adventurous I would say most people aren't adventurous in everything except their interests, which they are very adventurous in. The problem is that there will always be more people with a meh feeling toward a random subject than people with a genuine intrest in it, and a brute force ML algorithem trained by just throwing unholy amounts of generic personal data at it before being being relased on the giant mashup of random subjects that make up something like youtube won't catch that nuance
majormajor 8y ago
Whether product design or technical/mathematical, isn't that just saying that we have a problem of companies treating consumers as an aggregate userbase that they can train once against instead of everyone being a bit different?
eli_gottlieb 8y ago
>I don't think the problem is with ML. You could easily train a net to make boulder selections with an appropriate data set. Ehhh... mathematically, most ML methods basically assume that the underlying data and label manifolds should be smooth (in the real analysis sense). Sharply peaked output distributions, small mountains of "loved it!" surrounded by valleys of "ugh", are hard to work with for many models, compared to a wavy surface between "ok" and "ok, I guess".

AlexandrB 8y ago

>> We've positioned machine learning as a replacement to curation, but what it really does is put forth the least objectionable content. > I don't think the problem is with ML. You could easily train a net to make boulder selections with an appropriate data set. I believe what you are seeing is an accommodation for the average person. Most people aren't adventurous, and it makes business sense to get algos working for the middle majority than to cater to tail ends. You literally just said the same thing the GP did while ostensibly disagreeing with him: ML curated content caters to the lowest common denominator - i.e. the content that most people would be okay with.

Declanomous 8y ago

> boulder selections Would that mean more rock and less bluegrass? (I'm kidding, of course) In all seriousness though, the problem probably is the data set to a certain extent. I am pretty adventurous, true, but at the same time I think the problem is that regardless of what data we collect, we only have the ability to determine if something is inoffensive. All you need to do is look at a site that aggregates reviews like Yelp or Amazon, and read some reviews. 95% of reviews are either 1 star or 5 stars, and it's obvious from reading the text that the rating is meaningless. "I ordered the wrong item by mistake, 1 star" "Food is overpriced, but the bathrooms are clean and wait staff is attractive, 5 stars." I find star ratings most useful because people who rate things 2-4 stars generally have the most nuanced and productive things to say about the product, and actually rate things what they think they should be rated, rather than using the rating system as a vote for "rate this higher/lower." I don't think it's possible to actually come up with good recommendations based on user-reported like/dislike rating. It's not wrong for a user to dislike a song because it reminds them of an ex, but using that as a basis for a recommendation to someone else is entirely useless. Systems like Rotten Tomatoes works really well in this regard, but almost has the opposite problem, which is that it tends to underrate movies with broad appeal, but that's generally not a problem since users will be exposed to those movies anyways.

tomatotomato37 8y ago

>Most people aren't adventurous I would say most people aren't adventurous in everything except their interests, which they are very adventurous in. The problem is that there will always be more people with a meh feeling toward a random subject than people with a genuine intrest in it, and a brute force ML algorithem trained by just throwing unholy amounts of generic personal data at it before being being relased on the giant mashup of random subjects that make up something like youtube won't catch that nuance

majormajor 8y ago

Whether product design or technical/mathematical, isn't that just saying that we have a problem of companies treating consumers as an aggregate userbase that they can train once against instead of everyone being a bit different?

eli_gottlieb 8y ago

>I don't think the problem is with ML. You could easily train a net to make boulder selections with an appropriate data set. Ehhh... mathematically, most ML methods basically assume that the underlying data and label manifolds should be smooth (in the real analysis sense). Sharply peaked output distributions, small mountains of "loved it!" surrounded by valleys of "ugh", are hard to work with for many models, compared to a wavy surface between "ok" and "ok, I guess".