3 ms·

It depends on the definition of fairness, which is what I’m trying to argue for. Practically speaking, heavy correlation and casual relationships are similarly

by mirker 5y ago

It depends on the definition of fairness, which is what I’m trying to argue for. Practically speaking, heavy correlation and casual relationships are similarly problematic. If you correlate enough of the variables, you can limit your pool to a very particular group of people. And if your group of people is 99% of race/gender/age, then is that practically any different from doing a direct filter? The usual definitions of fairness actually have the strong constraint that the response (Y=f(X)) is independent of the fairness variables (f(X|Z) is the same for all fairness values of Z). So, concretely, if Z is gender, then you have the same insurance outcome regardless of gender. In this definition, the setting I described is unfair, since the correlation on education will bias the model to predict in a way that is correlated with race. The only practical way to satisfy such a strong definition is to do a correction on the data, which is a pretty straightforward optimization problem if you formalize the constraints. And yeah, usually you lose predictive power in doing so. For the record, the strong definition makes sense often times (to me). You wouldn’t want health insurance using your DNA as a predictor, likely, even if genetics are very predictive in that domain.

mattmcknight 5y ago
What depends on the definition of fairness? Surely not whether the queries are the same. The queries are only the same if one factor determines the other. This is why you immediately retreat to a 99%, which still is not practically the same, and still doesn't matter if the correct features are being captured. If I am looking to predict who can run a 4 minute mile, and it turns out my model returns no people over 60, is that unfair? Hardly. The record for an over 60 person is 4:49 so most good models would not return anyone over that age. However, if my model is based on various physical features, not including age, if a 61 year old shows up with those physical features, "age bias" will not stop the model from predicting them as capable. You are trying to shift the discussion to a 99% case, which is quite different from most correlations in the first place. This is clearly not the case for predictive variables like education. If you have a graduation rate of 90% for one group and 80% for another group, your algorithm is favoring some of each group, and disfavoring some of each group based on education, not group membership. > In this definition, the setting I described is unfair, since the correlation on education will bias the model to predict in a way that is correlated with race. You must be leaving something out of the explanation or I missed it. Say p(insurance_claim) = f(education). If this relationship holds across groups, the model is not biased. The model would be biased only if education were predictive for some subgroups and not others, which would point to education not actually being the cause, and another latent variable driving the relationship (say conscientiousness). You are suggesting that even if the relationship holds across groups, that education is predictive for all groups, that the model is still unfair, because education is not equal across all groups. I think this is more of a case of education perhaps not being fair or life not being fair than the model not being fair? If evidence of education, let's a say passing a driver safety course, is strongly, negatively correlated with insurance claims, but you find that some people cannot afford a driver safety course, it is not the model which is unfair, but perhaps the economic circumstances that led to someone not being able to afford the course. Let's not burden models of insurance risk or expected cost by making them less correct in order to pretend that unfairness in society does not exist. Having a model that is correlated with race is not evidence of bias in the model, but may be evidence of bias in society. > The only practical way to satisfy such a strong definition is to do a correction on the data, which is a pretty straightforward optimization problem if you formalize the constraints. And yeah, usually you lose predictive power in doing so. If, as in your example, you had a feature that was 99% determined by group membership, you would lose 99% of the value of that feature. This is why the "strong" definition is not helpful. > For the record, the strong definition makes sense often times (to me). You wouldn’t want health insurance using your DNA as a predictor, likely, even if genetics are very predictive in that domain. This is a completely different argument, that a more predictive model might not benefit you. However, it does get me to understand your viewpoint a little bit. You seem to want a world where the luck of what DNA you are born with does not affect you. However, if I were to use a model to predict my own healthcare costs using DNA and decide whether to buy long term care insurance, I certainly wouldn't want a model that was modified to remove "bias" and gave me a less accurate prediction of the real costs.
- mirker 5y ago
  Ok, a lot to unpack here. The viewpoint I am presenting is not novel and is what the paper is about. These “academic” notions of fairness are things you can prove for models, unlike ad-hoc variants that “feel” fair. So it’s not “my” viewpoint, but the broader “academic” viewpoint. If you look at section 5.2 of this paper and re-read my post, you will probably better understand what I mean https://arxiv.org/pdf/1912.06883.pdf https://arxiv.org/pdf/1912.06883.pdf A very similar example is discussed in 6.2 of https://arxiv.org/pdf/1104.3913.pdf https://arxiv.org/pdf/1104.3913.pdf, which is a seminal work in the field. Also see Appendix A point 2 for more discussion on encoding race/sex in correlations. My top-level post was asking what we care about, which is in-line with the running example. I doubt society really cares too much about fairness in running vs. say insurance claims and race. My point is that even with a perfect correction model, you still need to agree on what correction to apply. And yeah, for DNA, I agree with what you said. But if you use the same “unfair” model for non-personal use, you very quickly start running a eugenics program (e.g., deny healthcare for all people with a marker for cancer), which is generally viewed as bad. Basically, if you aren’t impacting society (e.g., personal use), then fairness is irrelevant.
  - mattmcknight 5y ago
    I am proposing a notion of fairness that you can prove. I am arguing that the other notion of fairness is not a useful measure of model fairness because it results in lower model performance and has negative real world consequences when employed. The example in 5.2 of the Binns paper suggests a definition of fair where the results of model have to be the same for different groups. None of the obvious problems are identified. If you take group A and B, with average income of 100,000 and 50,000 respectively, this fairness definition suggests that a loan approval model for Group A and B should have the same approval rate for a 50,000 loan in order to be fair. If this model were used, the model would be less accurate in predicting repayment than a model that did not attempt to correct for group disparities, Group B would default at a much higher rate and suffer the negative consequences to their credit. This notion of fairness denies the reality of the situation. If there is a group difference in ability to repay loans, it is not the job of the model to pretend that this difference does not exist and make decisions that will negatively affect the financial position of all concerned. The mention in 6.2 of the second paper didn't convey much on this topic, but they do admit "statistical parity is insufficient as a general notion of fairness." It does even take it back to Rawls, one philosopher I find most offensive. I was hopeful when they started with "We capture fairness by the principle that any two individuals who are similar with respect to a particular task should be classified similarly." Yet, they then dissolve this only having this condition hold within a subgroup, such that similar members of group A are treated similarly, and similar members of group B are treated similarly, but a member of group A that is similar to a member of group B is not treated similarly. The notion of fairness I advocate is that the selected features have to predictive, independent of group. Thus, the fairness measure for whether income could be included in the model would be whether income was predictive for group A and group B. If you took the data for group A and income was predictive and then then the data for group B and income was predictive, it would indicate that it was a valid measure for each groups. If you had a feature that was not predictive for each group, it would fail this fairness test and could not include it in an uncorrected model used for all groups. In terms of what we care about, I find it vastly better to discuss these topics with things we don't care much about. However, if we discover that driver safety courses really improve driver safety and lead to lower insurance costs, creating "fairness" by faking the results so that those who cannot afford these courses are given artificially lower insurance rates to compensate for this, we are actually being unfair to the society that would benefit from better drivers, rather than attempting to correct the underlying issue by subsidizing the cost of the course for drivers that cannot afford it.

by mirker 5y ago

mattmcknight 5y ago

What depends on the definition of fairness? Surely not whether the queries are the same. The queries are only the same if one factor determines the other. This is why you immediately retreat to a 99%, which still is not practically the same, and still doesn't matter if the correct features are being captured. If I am looking to predict who can run a 4 minute mile, and it turns out my model returns no people over 60, is that unfair? Hardly. The record for an over 60 person is 4:49 so most good models would not return anyone over that age. However, if my model is based on various physical features, not including age, if a 61 year old shows up with those physical features, "age bias" will not stop the model from predicting them as capable. You are trying to shift the discussion to a 99% case, which is quite different from most correlations in the first place. This is clearly not the case for predictive variables like education. If you have a graduation rate of 90% for one group and 80% for another group, your algorithm is favoring some of each group, and disfavoring some of each group based on education, not group membership. > In this definition, the setting I described is unfair, since the correlation on education will bias the model to predict in a way that is correlated with race. You must be leaving something out of the explanation or I missed it. Say p(insurance_claim) = f(education). If this relationship holds across groups, the model is not biased. The model would be biased only if education were predictive for some subgroups and not others, which would point to education not actually being the cause, and another latent variable driving the relationship (say conscientiousness). You are suggesting that even if the relationship holds across groups, that education is predictive for all groups, that the model is still unfair, because education is not equal across all groups. I think this is more of a case of education perhaps not being fair or life not being fair than the model not being fair? If evidence of education, let's a say passing a driver safety course, is strongly, negatively correlated with insurance claims, but you find that some people cannot afford a driver safety course, it is not the model which is unfair, but perhaps the economic circumstances that led to someone not being able to afford the course. Let's not burden models of insurance risk or expected cost by making them less correct in order to pretend that unfairness in society does not exist. Having a model that is correlated with race is not evidence of bias in the model, but may be evidence of bias in society. > The only practical way to satisfy such a strong definition is to do a correction on the data, which is a pretty straightforward optimization problem if you formalize the constraints. And yeah, usually you lose predictive power in doing so. If, as in your example, you had a feature that was 99% determined by group membership, you would lose 99% of the value of that feature. This is why the "strong" definition is not helpful. > For the record, the strong definition makes sense often times (to me). You wouldn’t want health insurance using your DNA as a predictor, likely, even if genetics are very predictive in that domain. This is a completely different argument, that a more predictive model might not benefit you. However, it does get me to understand your viewpoint a little bit. You seem to want a world where the luck of what DNA you are born with does not affect you. However, if I were to use a model to predict my own healthcare costs using DNA and decide whether to buy long term care insurance, I certainly wouldn't want a model that was modified to remove "bias" and gave me a less accurate prediction of the real costs.

mirker 5y ago
Ok, a lot to unpack here. The viewpoint I am presenting is not novel and is what the paper is about. These “academic” notions of fairness are things you can prove for models, unlike ad-hoc variants that “feel” fair. So it’s not “my” viewpoint, but the broader “academic” viewpoint. If you look at section 5.2 of this paper and re-read my post, you will probably better understand what I mean https://arxiv.org/pdf/1912.06883.pdf https://arxiv.org/pdf/1912.06883.pdf A very similar example is discussed in 6.2 of https://arxiv.org/pdf/1104.3913.pdf https://arxiv.org/pdf/1104.3913.pdf, which is a seminal work in the field. Also see Appendix A point 2 for more discussion on encoding race/sex in correlations. My top-level post was asking what we care about, which is in-line with the running example. I doubt society really cares too much about fairness in running vs. say insurance claims and race. My point is that even with a perfect correction model, you still need to agree on what correction to apply. And yeah, for DNA, I agree with what you said. But if you use the same “unfair” model for non-personal use, you very quickly start running a eugenics program (e.g., deny healthcare for all people with a marker for cancer), which is generally viewed as bad. Basically, if you aren’t impacting society (e.g., personal use), then fairness is irrelevant.
- mattmcknight 5y ago
  I am proposing a notion of fairness that you can prove. I am arguing that the other notion of fairness is not a useful measure of model fairness because it results in lower model performance and has negative real world consequences when employed. The example in 5.2 of the Binns paper suggests a definition of fair where the results of model have to be the same for different groups. None of the obvious problems are identified. If you take group A and B, with average income of 100,000 and 50,000 respectively, this fairness definition suggests that a loan approval model for Group A and B should have the same approval rate for a 50,000 loan in order to be fair. If this model were used, the model would be less accurate in predicting repayment than a model that did not attempt to correct for group disparities, Group B would default at a much higher rate and suffer the negative consequences to their credit. This notion of fairness denies the reality of the situation. If there is a group difference in ability to repay loans, it is not the job of the model to pretend that this difference does not exist and make decisions that will negatively affect the financial position of all concerned. The mention in 6.2 of the second paper didn't convey much on this topic, but they do admit "statistical parity is insufficient as a general notion of fairness." It does even take it back to Rawls, one philosopher I find most offensive. I was hopeful when they started with "We capture fairness by the principle that any two individuals who are similar with respect to a particular task should be classified similarly." Yet, they then dissolve this only having this condition hold within a subgroup, such that similar members of group A are treated similarly, and similar members of group B are treated similarly, but a member of group A that is similar to a member of group B is not treated similarly. The notion of fairness I advocate is that the selected features have to predictive, independent of group. Thus, the fairness measure for whether income could be included in the model would be whether income was predictive for group A and group B. If you took the data for group A and income was predictive and then then the data for group B and income was predictive, it would indicate that it was a valid measure for each groups. If you had a feature that was not predictive for each group, it would fail this fairness test and could not include it in an uncorrected model used for all groups. In terms of what we care about, I find it vastly better to discuss these topics with things we don't care much about. However, if we discover that driver safety courses really improve driver safety and lead to lower insurance costs, creating "fairness" by faking the results so that those who cannot afford these courses are given artificially lower insurance rates to compensate for this, we are actually being unfair to the society that would benefit from better drivers, rather than attempting to correct the underlying issue by subsidizing the cost of the course for drivers that cannot afford it.