3 ms·

Someone should make a version of reddit where your upvote/downvote record is what shapes the ordering for you. There should be a global rank and an individual

Someone should make a version of reddit where your upvote/downvote record is what shapes the ordering for you. There should be a global rank and an individual rank that is calculated for each user something like the netflix 5 star system. That way the hateful stuff won't appear unless you upvoted hateful stuff. Liberals and conservatives could both get the content they want going to the same politics forum. It should be assumed that when the professional executive class gets their hands on the company they will try to arbitrarily censor and ban things. The system has to be decentralized so the parent company is technically unable to do this even if north korea were to purchase it.

bmh100 11y ago
What you are essentially describing, also used by Netflix, is a "collaborative filtering" machine learning algorithm [1]. It is a fascinating sparse-matrix problem that could be applied to many situations dependent upon individual preferences. [1]: https://en.wikipedia.org/wiki/Collaborative_filtering https://en.wikipedia.org/wiki/Collaborative_filtering
- c_prompt 11y ago
  For my site, I coded a recommendation engine based on cosine similarity. I initially tried Bayes' Theorem and conditional probabilities but abandoned the effort when I didn't see results that made a lot of sense (probably due to a lack of training data). Instead, I used https://github.com/aoiaoi/CosineSimilarity https://github.com/aoiaoi/CosineSimilarity as a base and then modified. If the results prove effective over time, I'm going to make the algorithm an option for users to select for generating their front page instead of upvotes.
drdeca 11y ago
Sounds a little bit similar to this one idea I read about, but am having trouble finding right now (fakeedit before submit: it was https://github.com/neyer/respect https://github.com/neyer/respect ), which was meant to be a system for determining respect and reputation in large groups. The idea was that each person would select some number of people to state that they trust/respect (fakeedit: this should say respect most of the times I say trust, but I don't feel like replacing all of them atm. mentally substitute the word 'trust' with 'respect') them to some degree (between 0 and 1, or between -1 and 1, but the negative ones were counted as zero until the last step). These would together form a (square) matrix of how much each person has granted 'initial' trust to who. Also each person's statement of who they trust is a vector, and it is normalized to have magnitude 1 (before being used as part of the matrix. This is to defend against attacks from people making extra identities that all trust each other, or things like that) The idea would then be that people tend to trust people that people they trust, trust, so it would be useful to determine what trust levels there are when including 'indirect' trust. To do this, one adds the matrix to the matrix raised to the second power, plus the matrix raised to the third power, for some number of levels (I think they recommended 3 levels). For each of these, on the last level, one can include the negative values in order to include /distrust/. (But one does not want to include it in earlier levels, otherwise one would end up trusting people that people you distrust, distrust, which probably isn't what you want). Then to find someone's indirect trust vector you look at the appropriate, uh, I think row, of the resulting matrix. In addition, there is a measure of how much your 'initial' trust vector aligns with your 'indirect' trust vector, by taking the dot product of the two, (and I think also dividing by the magnitudes of the two vectors). This is sort of a measure of how similar your trust is to that of the people you trust. .... oh, I think I originally saw it as a blog post, but it looks like it was by neyer and is at https://github.com/neyer/respect https://github.com/neyer/respect ... So, a roughly analogue thing for using upvotes as the input would I guess be, uh... To determine the ranking that a given person would see, consider the list of all of a person's upvotes/downvotes (or all in the last 6 months or w/e?) as a vector, and for each person who upvoted or downvoted anything in the thread one is looking at, take the dot product of their voting record vector with one's own voting record, and weigh their votes in the given thread based on that dot product (perhaps adjusted for the magnitude of the voting record vectors in question). So, if someone is someone who you vote similarly to, their upvotes increase the score you see of posts, and if you tend to vote the opposite way from them, their upvotes would decrease the score you see on posts (the opposite would be the case with downvotes). However, this seems like it might be computationally expensive? For each thread one looks at, getting each person who voted a single time in that thread, and then get every time they voted in the last <x> time ( or at least all the times that they voted, where you also voted), and compare all of your (recent) votes with theirs. Another concern that I think people would have would be that it might produce some echo-chamber-ish effects, but I don't know if that would really be a problem. It might be that if one votes for what makes good arguments for their position, and not just the opinions that one agrees with, then one will vote similarly to other people who do that, and that one will as a result see the posts that people who do that upvoted. But if there is a group who does not do that, and is large, and many of them vote for a post which the people who vote for good reasoning, that they might have an unfortunate influence on how the people who are voting for careful arguments see. However, it seems like this would be solved by simply downvoting the arguments that do not contribute. So, depending on how one defines "echo chamber", it seems like the system might only give one an "echo chamber" if one votes in a way that would align with that. However, for another seemingly reasonable definition of "echo chamber" , I guess it would lead to one (with that being essentially the purpose), but the fact that it does so might not be all that bad (with the people who are "open minded" ending up in an "echo chamber" of other "open minded" people). It also seems like it could be combined with github:neyer/respect in potentially nice ways. for example, the dot products of similarity of voting seem to be possibly of the same "type" as the respect value in github:neyer/respect , and it seems like maybe they could be added together. So, one could maybe do something like add the indirect respect value to the similarity of voting value, when weighing how much to consider the different votes, or perhaps one could use the dot products of the voting records as contributing to the initial trust vectors. (In either case, this could be potentially weighted to make one have more weight than the other). This might make it so that one has more people in the vector of who one initially has respect information about, which could be useful. But I don't know if similarity in voting record is a good thing for that. And there's still the problem of whether it is too computationally expensive to do for every thread view for every person. Maybe if there was some caching, and some of the stuff didn't get updated all the time? But that seems like it would require a lot of memory. _____ another potential problem is, if it is to be decentralized, it would I think be difficult to use the above system without making lists of who upvoted/downvoted who entirely public (unless one used something like the recent MIT Enigma thing, but that would make this all even more computationally and memory expensive. /Mayyyybe/ not prohibitively so though? I'm not sure. Practicality isn't my strong suit.) _____ It is a very interesting problem!
meesterdude 11y ago
> Liberals and conservatives could both get the content they want going to the same politics forum. Hey, that's my thought! And I thought I was being original. Also probably not the first to think that either. The whole "shapes the ordering for you" idea isn't bad; I think maybe even is a good one, largely. But, the one thing I take issue with is customization, or being more mosaic. Movies and TV shows it totally makes sense for; but I think it's a little more nuanced with news sites given the broader scope of topics and such.
renata 11y ago
Amusingly, once upon a time, this was reddit's goal. Everyone had a "Recommended" tab, and it never ever worked. They killed it years ago.

bmh100 11y ago

What you are essentially describing, also used by Netflix, is a "collaborative filtering" machine learning algorithm [1]. It is a fascinating sparse-matrix problem that could be applied to many situations dependent upon individual preferences. [1]: https://en.wikipedia.org/wiki/Collaborative_filtering https://en.wikipedia.org/wiki/Collaborative_filtering

c_prompt 11y ago
For my site, I coded a recommendation engine based on cosine similarity. I initially tried Bayes' Theorem and conditional probabilities but abandoned the effort when I didn't see results that made a lot of sense (probably due to a lack of training data). Instead, I used https://github.com/aoiaoi/CosineSimilarity https://github.com/aoiaoi/CosineSimilarity as a base and then modified. If the results prove effective over time, I'm going to make the algorithm an option for users to select for generating their front page instead of upvotes.

drdeca 11y ago

Sounds a little bit similar to this one idea I read about, but am having trouble finding right now (fakeedit before submit: it was https://github.com/neyer/respect https://github.com/neyer/respect ), which was meant to be a system for determining respect and reputation in large groups. The idea was that each person would select some number of people to state that they trust/respect (fakeedit: this should say respect most of the times I say trust, but I don't feel like replacing all of them atm. mentally substitute the word 'trust' with 'respect') them to some degree (between 0 and 1, or between -1 and 1, but the negative ones were counted as zero until the last step). These would together form a (square) matrix of how much each person has granted 'initial' trust to who. Also each person's statement of who they trust is a vector, and it is normalized to have magnitude 1 (before being used as part of the matrix. This is to defend against attacks from people making extra identities that all trust each other, or things like that) The idea would then be that people tend to trust people that people they trust, trust, so it would be useful to determine what trust levels there are when including 'indirect' trust. To do this, one adds the matrix to the matrix raised to the second power, plus the matrix raised to the third power, for some number of levels (I think they recommended 3 levels). For each of these, on the last level, one can include the negative values in order to include /distrust/. (But one does not want to include it in earlier levels, otherwise one would end up trusting people that people you distrust, distrust, which probably isn't what you want). Then to find someone's indirect trust vector you look at the appropriate, uh, I think row, of the resulting matrix. In addition, there is a measure of how much your 'initial' trust vector aligns with your 'indirect' trust vector, by taking the dot product of the two, (and I think also dividing by the magnitudes of the two vectors). This is sort of a measure of how similar your trust is to that of the people you trust. .... oh, I think I originally saw it as a blog post, but it looks like it was by neyer and is at https://github.com/neyer/respect https://github.com/neyer/respect ... So, a roughly analogue thing for using upvotes as the input would I guess be, uh... To determine the ranking that a given person would see, consider the list of all of a person's upvotes/downvotes (or all in the last 6 months or w/e?) as a vector, and for each person who upvoted or downvoted anything in the thread one is looking at, take the dot product of their voting record vector with one's own voting record, and weigh their votes in the given thread based on that dot product (perhaps adjusted for the magnitude of the voting record vectors in question). So, if someone is someone who you vote similarly to, their upvotes increase the score you see of posts, and if you tend to vote the opposite way from them, their upvotes would decrease the score you see on posts (the opposite would be the case with downvotes). However, this seems like it might be computationally expensive? For each thread one looks at, getting each person who voted a single time in that thread, and then get every time they voted in the last <x> time ( or at least all the times that they voted, where you also voted), and compare all of your (recent) votes with theirs. Another concern that I think people would have would be that it might produce some echo-chamber-ish effects, but I don't know if that would really be a problem. It might be that if one votes for what makes good arguments for their position, and not just the opinions that one agrees with, then one will vote similarly to other people who do that, and that one will as a result see the posts that people who do that upvoted. But if there is a group who does not do that, and is large, and many of them vote for a post which the people who vote for good reasoning, that they might have an unfortunate influence on how the people who are voting for careful arguments see. However, it seems like this would be solved by simply downvoting the arguments that do not contribute. So, depending on how one defines "echo chamber", it seems like the system might only give one an "echo chamber" if one votes in a way that would align with that. However, for another seemingly reasonable definition of "echo chamber" , I guess it would lead to one (with that being essentially the purpose), but the fact that it does so might not be all that bad (with the people who are "open minded" ending up in an "echo chamber" of other "open minded" people). It also seems like it could be combined with github:neyer/respect in potentially nice ways. for example, the dot products of similarity of voting seem to be possibly of the same "type" as the respect value in github:neyer/respect , and it seems like maybe they could be added together. So, one could maybe do something like add the indirect respect value to the similarity of voting value, when weighing how much to consider the different votes, or perhaps one could use the dot products of the voting records as contributing to the initial trust vectors. (In either case, this could be potentially weighted to make one have more weight than the other). This might make it so that one has more people in the vector of who one initially has respect information about, which could be useful. But I don't know if similarity in voting record is a good thing for that. And there's still the problem of whether it is too computationally expensive to do for every thread view for every person. Maybe if there was some caching, and some of the stuff didn't get updated all the time? But that seems like it would require a lot of memory. _____ another potential problem is, if it is to be decentralized, it would I think be difficult to use the above system without making lists of who upvoted/downvoted who entirely public (unless one used something like the recent MIT Enigma thing, but that would make this all even more computationally and memory expensive. /Mayyyybe/ not prohibitively so though? I'm not sure. Practicality isn't my strong suit.) _____ It is a very interesting problem!

meesterdude 11y ago

> Liberals and conservatives could both get the content they want going to the same politics forum. Hey, that's my thought! And I thought I was being original. Also probably not the first to think that either. The whole "shapes the ordering for you" idea isn't bad; I think maybe even is a good one, largely. But, the one thing I take issue with is customization, or being more mosaic. Movies and TV shows it totally makes sense for; but I think it's a little more nuanced with news sites given the broader scope of topics and such.

renata 11y ago

Amusingly, once upon a time, this was reddit's goal. Everyone had a "Recommended" tab, and it never ever worked. They killed it years ago.