5 ms·

Data that's had data censored from it is bad data.

dredmorbius 5y ago
False. Outlier exclusion is standard practice. It's useful to note what is excluded. But you exclude bad data from the analysis. Remember that what you're interested in is not the data but the ground truth that the data represent. This means that the full transmission chain must be reliable and its integrity assured: phenomenon, generated signal, transmission channel, receiver, sensor, interpretation, and recording. Noise may enter at any point. And that noise has ... exceedingly little value. Deliberately inserted noise is one of the most effective ways to thwart an accurate assessment of ground truths.
ineedasername 5y ago
Defining terms here is important, so let's avoid the word bad for a moment because it can be applied in different ways. 1) You can have an empty dataset. 2) You can have an incomplete dataset. 3) you can have a dataset where the data is wrong All of these situations, in some sense, are "bad" What I'm saying is that, going into a situation, my preference would be #2 > #1 > #3. Because I always assume a dataset could be incomplete, that it didn't capture everything. I can plan for it, look for evidence that something is missing, try to find it. If I suspect something is missing but can't find it then I at least know that much, and maybe even the magnitude of uncertainty that adds to the situation. Either way, I can work around it understanding the limits if what I'm doing or if there's too much missing, make a judgement call and say that nothing useful can be done with it. If I have what appears to be a dataset that I can work with, but the data is all incorrect, I may never even know it until things start to break or, before that if I'm lucky, I waste large amounts of time to find out that the results just don't make sense. It's probably important to note that #2 and #3 are also not mutually exclusive. Getting out of the dry world of data analysis, if your job is propaganda & if you're good at your job, #2 and #3 combined is where you're at.
- inkblotuniverse 5y ago
  I'd argue Facebook's censorship leaves us with 2 and 3. They don't remove things bevause they're wrong; they remove them because they go against the current orthodoxy. Most things are wrong, so most things that go against the modern orthodoxy are wrong... but wrong things that go WITH the modern orthodoxy aren't removed. It's a scientist who removes outliers in the direction that refute his ideas, but not ones in the direction that support it.
  - dredmorbius 5y ago
    Let's note that this thread's been shifting back and forth between information which is publicised over media and data, with the discussion focusing on use in research. These aren't entirely dissimilar, but they have both similarities and differences. Data in research is used to confirm or deny models, that is, understandings of the world. Data in operations is used to determine and shape actions (including possibly inaction), interacting with an environment. Information in media ... shares some of this, but is more complex in that it both creates (or disproves) models, and has a very extensive behavioural component involving both individual and group psychology and sociology. Media platform moderation plays several roles. In part, it's performed in the context that the platforms are performing their own selection and amplification, and that there's now experimental evidence that even in the absence of any induced bias, disinformation tends to spread especially in large and active social networks. (See "Information Overload Helps Fake News Spread, and Social Media Knows It". (https://www.scientificamerican.com/article/information-overload-helps-fake-news-spread-and-social-media-knows-it/ https://www.scientificamerican.com/article/information-overl...), discussed here https://news.ycombinator.com/item?id=28495912 https://news.ycombinator.com/item?id=28495912 and https://news.ycombinator.com/item?id=25153716 https://news.ycombinator.com/item?id=25153716) The situation is made worse when there's both intrinsic tooling of the system to boost sensationalism (a/k/a "high engagement" content), and deliberate introduction of false or provocative information. TL;DR: moderation has to compensate and overcome inherent biases for misinformation, and take into consideration both causal and resultant behaviours and effects. At the same time, moderation itself is subject to many of the same biases that the information network as a whole is (false and inflammatory reports tend to draw more reports and quicker actions), as well as spurious error rates (as I've described at length above). All of which is to say that I don't find your own allegation of an intentional bias, offered without evidence or argument, credible.

False. Outlier exclusion is standard practice. It's useful to note what is excluded. But you exclude bad data from the analysis. Remember that what you're interested in is not the data but the ground truth that the data represent. This means that the full transmission chain must be reliable and its integrity assured: phenomenon, generated signal, transmission channel, receiver, sensor, interpretation, and recording. Noise may enter at any point. And that noise has ... exceedingly little value. Deliberately inserted noise is one of the most effective ways to thwart an accurate assessment of ground truths.

ineedasername 5y ago

Defining terms here is important, so let's avoid the word bad for a moment because it can be applied in different ways. 1) You can have an empty dataset. 2) You can have an incomplete dataset. 3) you can have a dataset where the data is wrong All of these situations, in some sense, are "bad" What I'm saying is that, going into a situation, my preference would be #2 > #1 > #3. Because I always assume a dataset could be incomplete, that it didn't capture everything. I can plan for it, look for evidence that something is missing, try to find it. If I suspect something is missing but can't find it then I at least know that much, and maybe even the magnitude of uncertainty that adds to the situation. Either way, I can work around it understanding the limits if what I'm doing or if there's too much missing, make a judgement call and say that nothing useful can be done with it. If I have what appears to be a dataset that I can work with, but the data is all incorrect, I may never even know it until things start to break or, before that if I'm lucky, I waste large amounts of time to find out that the results just don't make sense. It's probably important to note that #2 and #3 are also not mutually exclusive. Getting out of the dry world of data analysis, if your job is propaganda & if you're good at your job, #2 and #3 combined is where you're at.

inkblotuniverse 5y ago
I'd argue Facebook's censorship leaves us with 2 and 3. They don't remove things bevause they're wrong; they remove them because they go against the current orthodoxy. Most things are wrong, so most things that go against the modern orthodoxy are wrong... but wrong things that go WITH the modern orthodoxy aren't removed. It's a scientist who removes outliers in the direction that refute his ideas, but not ones in the direction that support it.
- dredmorbius 5y ago
  Let's note that this thread's been shifting back and forth between information which is publicised over media and data, with the discussion focusing on use in research. These aren't entirely dissimilar, but they have both similarities and differences. Data in research is used to confirm or deny models, that is, understandings of the world. Data in operations is used to determine and shape actions (including possibly inaction), interacting with an environment. Information in media ... shares some of this, but is more complex in that it both creates (or disproves) models, and has a very extensive behavioural component involving both individual and group psychology and sociology. Media platform moderation plays several roles. In part, it's performed in the context that the platforms are performing their own selection and amplification, and that there's now experimental evidence that even in the absence of any induced bias, disinformation tends to spread especially in large and active social networks. (See "Information Overload Helps Fake News Spread, and Social Media Knows It". (https://www.scientificamerican.com/article/information-overload-helps-fake-news-spread-and-social-media-knows-it/ https://www.scientificamerican.com/article/information-overl...), discussed here https://news.ycombinator.com/item?id=28495912 https://news.ycombinator.com/item?id=28495912 and https://news.ycombinator.com/item?id=25153716 https://news.ycombinator.com/item?id=25153716) The situation is made worse when there's both intrinsic tooling of the system to boost sensationalism (a/k/a "high engagement" content), and deliberate introduction of false or provocative information. TL;DR: moderation has to compensate and overcome inherent biases for misinformation, and take into consideration both causal and resultant behaviours and effects. At the same time, moderation itself is subject to many of the same biases that the information network as a whole is (false and inflammatory reports tend to draw more reports and quicker actions), as well as spurious error rates (as I've described at length above). All of which is to say that I don't find your own allegation of an intentional bias, offered without evidence or argument, credible.