4 ms·

The algorithms seem to favor popularity instead of discovering new options. Anything where you have followers/reviews will favor established players.

vannevar 5y ago
This is true of human nature as well, not just the search algorithms. As long as people have even a slight preference for the more popular option, you have the basic ingredient for a rich-get-richer market with power law distribution of sales.
s1artibartfast 5y ago
Put differently, algorithms favor satisfying viewer over discovering new options. This favors the established players, but this is because the average viewer is more likely to be satisfied with the most popular option than a random option.
- derbOac 5y ago
  I don't know. There's been randomized controlled experiments showing that providing information on popularity distorts choice to favor those who are already popular, so it's not just that people refer the more popular options. What you say is true, but it's complicated because most viewers are not the average viewer, by definition, so providing them with the average viewer preference could lead to most viewers being less satisfied. I do think there's some kind of feedback loop that plays into this system. I don't think it's unique to social networks or new but I do think the process of matching any kind of consumer with content will run into this. It's more of a regression to the mean problem in prediction (which is maybe what you're saying I suppose). I have this weird experience when I try to go to YouTube in a temporary container in Firefox without realizing it, and I'm served what the average visitor would see. It's so different from my usual homepage it's like a different world. Even so, I seem to run into these local minima in recommendations where I end up getting recommended the same stuff over and over and over again. Eventually I'll learn about some content from someplace else and it breaks out of that local minima and starts spinning new recommendations, but it's like I know there's other stuff out there I'd like, but it's not getting to me because the recommendation system only has information on things it's collected information on, which means things that have been viewed, which means more popular content. That's not even getting into issues like bands and artists being "rediscovered" decades after they were producing most of their music. Sometimes I feel as if these discussions devolve into this tug between "less popular is less popular because it's lower quality" versus "less popular is less popular because it doesn't get visibility" when the truth is in between. I suspect there's a lot of content out there that could get more exposure but doesn't because there's no match with the right audience, but if you're not really appealing to many people you'll never get anywhere. But then again, people change, and what's popular changes, and changes differently among different people... I think it's a really complicated issue involving feedback loops, discovery, information, and so forth. It's not just like finding some static gold needle in the haystack, it's like if the perception of what a gold needle is changes from person to person, and from time to time for the same person, and seeing another needle also changes that perception. How this maps onto the distribution of reimbursement? Not sure, but I suspect there's a disconnect between some ideal optimum situation and what's out there.
- stubish 5y ago
  I don't think this is quite true. I posit that only a minority of people will be satisfied with the most popular option when there are a large number of options. If your category is 'all viewers', the most popular choice will be poor for the majority, because 'all viewers' is too diverse. You need better categorization of the viewers before you can meaningfully suggest content beyond 'shovel more of this profitable content'. It is starting to happen with the 'people who liked X also liked Y' categories, but is still pretty poor (eg. recommending content a user hates without giving the user a way to teach the system that they hate that content or why they hate it). Machine learning would be a nice solution, but doesn't seem to scale to large numbers of users.

vannevar 5y ago

This is true of human nature as well, not just the search algorithms. As long as people have even a slight preference for the more popular option, you have the basic ingredient for a rich-get-richer market with power law distribution of sales.

s1artibartfast 5y ago

Put differently, algorithms favor satisfying viewer over discovering new options. This favors the established players, but this is because the average viewer is more likely to be satisfied with the most popular option than a random option.

derbOac 5y ago
I don't know. There's been randomized controlled experiments showing that providing information on popularity distorts choice to favor those who are already popular, so it's not just that people refer the more popular options. What you say is true, but it's complicated because most viewers are not the average viewer, by definition, so providing them with the average viewer preference could lead to most viewers being less satisfied. I do think there's some kind of feedback loop that plays into this system. I don't think it's unique to social networks or new but I do think the process of matching any kind of consumer with content will run into this. It's more of a regression to the mean problem in prediction (which is maybe what you're saying I suppose). I have this weird experience when I try to go to YouTube in a temporary container in Firefox without realizing it, and I'm served what the average visitor would see. It's so different from my usual homepage it's like a different world. Even so, I seem to run into these local minima in recommendations where I end up getting recommended the same stuff over and over and over again. Eventually I'll learn about some content from someplace else and it breaks out of that local minima and starts spinning new recommendations, but it's like I know there's other stuff out there I'd like, but it's not getting to me because the recommendation system only has information on things it's collected information on, which means things that have been viewed, which means more popular content. That's not even getting into issues like bands and artists being "rediscovered" decades after they were producing most of their music. Sometimes I feel as if these discussions devolve into this tug between "less popular is less popular because it's lower quality" versus "less popular is less popular because it doesn't get visibility" when the truth is in between. I suspect there's a lot of content out there that could get more exposure but doesn't because there's no match with the right audience, but if you're not really appealing to many people you'll never get anywhere. But then again, people change, and what's popular changes, and changes differently among different people... I think it's a really complicated issue involving feedback loops, discovery, information, and so forth. It's not just like finding some static gold needle in the haystack, it's like if the perception of what a gold needle is changes from person to person, and from time to time for the same person, and seeing another needle also changes that perception. How this maps onto the distribution of reimbursement? Not sure, but I suspect there's a disconnect between some ideal optimum situation and what's out there.
stubish 5y ago
I don't think this is quite true. I posit that only a minority of people will be satisfied with the most popular option when there are a large number of options. If your category is 'all viewers', the most popular choice will be poor for the majority, because 'all viewers' is too diverse. You need better categorization of the viewers before you can meaningfully suggest content beyond 'shovel more of this profitable content'. It is starting to happen with the 'people who liked X also liked Y' categories, but is still pretty poor (eg. recommending content a user hates without giving the user a way to teach the system that they hate that content or why they hate it). Machine learning would be a nice solution, but doesn't seem to scale to large numbers of users.