Digifesto

Tag: fairness in machine learning

All the problems with our paper, “Racial categories in machine learning”

Bruce Haynes and I were blown away by the reception to our paper, “Racial categories in machine learning“. This was a huge experiment in interdisciplinary collaboration for us. We are excited about the next steps in this line of research.

That includes engaging with criticism. One of our goals was to fuel a conversation in the research community about the operationalization of race. That isn’t a question that can be addressed by any one paper or team of researchers. So one thing we got out of the conference was great critical feedback on potential problems with the approach we proposed.

This post is an attempt to capture those critiques.

Need for participatory design

Khadijah Abdurahman, of Word to RI , issues a subtweeted challenge to us to present our paper to the hood. (RI stands for Roosevelt Island, in New York City, the location of the recently established Cornell Tech campus.)

On striking challenge, raised by Khadijah Abdurahman on Twitter, is that we should be developing peer relationships with the communities we research. I read this as a call for participatory design. It’s true this was not part of the process of the paper. In particular, Ms. Abdurahman points to a part of our abstract that uses jargon from computer science.

There are a lot of ways to respond to this comment. The first is to accept the challenge. I would personally love it if Bruce and I could present our research to folks on Roosevelt Island and get feedback from them.

There are other ways to respond that address the tensions of this comment. One is to point out that in addition to being an accomplished scholar of the sociology of race and how it forms, especially in urban settings, Bruce is a black man who is originally from Harlem. Indeed, Bruce’s family memoir shows his deep and well-researched familiarity with the life of marginalized people of the hood. So a “peer relationship” between an algorithm designer (me) and a member of an affected community (Bruce) is really part of the origin of our work.

Another is to point out that we did not research a particular community. Our paper was not human subjects research; it was about the racial categories that are maintained by the Federal U.S. government and which pervade society in a very general way. Indeed, everybody is affected by these categories. When I and others who looks like me are ascribed “white”, that is an example of these categories at work. Bruce and I were very aware of how different kinds of people at the conference responded to our work, and how it was an intervention in our own community, which is of course affected by these racial categories.

The last point is that computer science jargon is alienating to basically everybody who is not trained in computer science, whether they live in the hood or not. And the fact is we presented our work at a computer science venue. Personally, I’m in favor of universal education in computational statistics, but that is a tall order. If our work becomes successful, I could see it becoming part of, for example, a statistical demography curriculum that could be of popular interest. But this is early days.

The Quasi-Racial (QR) Categories are Not Interpretable

In our presentation, we introduced some terminology that did not make it into the paper. We named the vectors of segregation derived by our procedure “quasi-racial” (QR) vectors, to denote that we were trying to capture dimensions that were race-like, in that they captured the patterns of historic and ongoing racial injustice, without being the racial categories themselves, which we argued are inherently unfair categories of inequality.

First, we are not wedded to the name “quasi-racial” and are very open to different terminology if anybody has an idea for something better to call them.

More importantly, somebody pointed out that these QR vectors may not be interpretable. Given that the conference is not only about Fairness, but also Accountability and Transparency, this critique is certainly on point.

To be honest, I have not yet done the work of surveying the extensive literature on algorithm interpretability to get a nuanced response. I can give two informal responses. The first is that one assumption of our proposal is that there is something wrong with how race and racial categories are intuitive understood. Normal people’s understanding of race is, of course, ridden with stereotypes, implicit biases, false causal models, and so on. If we proposed an algorithm that was fully “interpretable” according to most people’s understanding of what race is, that algorithm would likely have racist or racially unequal outcomes. That’s precisely the problem that we are trying to get at with our work. In other words, when categories are inherently unfair, interpretability and fairness may be at odds.

The second response is that educating people about how the procedure works and why its motivated is part of what makes its outcomes interpretable. Teaching people about the history of racial categories, and how those categories are both the cause and effect of segregation in space and society, makes the algorithm interpretable. Teaching people about Principal Component Analysis, the algorithm we employ, is part of what makes the system interpretable. We are trying to drop knowledge; I don’t think we are offering any shortcuts.

Principal Component Analysis (PCA) may not be the right technique

An objection from the computer science end of the spectrum was that our proposed use of Principal Component Analysis (PCA) was not well-motivated enough. PCA is just one of many dimensionality reduction techniques–why did we choose it in particular? PCA has many assumptions about the input embedded within it, including the component vectors of interest are linear combinations of the inputs. What if the best QR representation is a non-linear combination of the input variables? And our use of unsupervised learning, as a general criticism, is perhaps lazy, since in order to validate its usefulness we will need to test it with labeled data anyway. We might be better off with a more carefully calibrated and better motivated alternative technique.

These are all fair criticisms. I am personally not satisfied with the technical component of the paper and presentation. I know the rigor of the analysis is not of the standard that would impress a machine learning scholar and can take full responsibility for that. I hope to do better in a future iteration of the work, and welcome any advice on how to do that from colleagues. I’d also be interested to see how more technically skilled computer scientists and formal modelers address the problem of unfair racial categories that we raised in the paper.

I see our main contribution as the raising of this problem of unfair categories, not our particular technical solution to it. As a potential solution, I hope that it’s better than nothing, a step in the right direction, and provocative. I subscribe to the belief that science is an iterative process and look forward to the next cycle of work.

Please feel free to reach out if you have a critique of our work that we’ve missed. We do appreciate all the feedback!

Advertisements

Notes on O’Neil, Chapter 2, “Bomb Parts”

Continuing with O’Neil’s Weapons of Math Destruction on to Chapter 2, “Bomb Parts”. This is a popular book and these are quick chapters. But that’s no reason to underestimate them! This is some of the most lucid work I’ve read on algorithmic fairness.

This chapter talks about three kinds of “models” used in prediction and decision making, with three examples. O’Neil speak highly of the kinds of models used in baseball to predict the trajectory of hits and determine the optimal placement of people in the field. (Ok, I’m not so good at baseball terms). These are good, O’Neil says, because they are transparent, they are consistently adjusted with new data, and the goals are well defined.

O’Neil then very charmingly writes about the model she uses mentally to determine how to feed her family. She juggles a lot of variables: the preferences of her kids, the nutrition and cost of ingredients, and time. This is all hugely relatable–everybody does something like this. Her point, it seems, is that this form of “model” encodes a lot of opinions or “ideology” because it reflects her values.

O’Neil then discusses recidivism prediction, specifically the LSI-R (Level of Service Inventory–Revised) tool. It asks questions like “How many previous convictions have you had?” and uses that to predict likelihood of future prediction. The problem is that (a) this is sensitive to overpolicing in neighborhoods, which has little to do with actual recidivism rates (as opposed to rearrest rates), and (b) e.g. black neighborhoods are more likely to be overpoliced, meaning that the tool, which is not very good at predicting recidivism, has disparate impact. This is an example of what O’Neil calls an (eponymous) weapon of math destruction.(WMD)

She argues that the three qualities of a WMD are Scale, Opacity, and Damage. Which makes sense.

As I’ve said, I think this is a better take on algorithmic ethics than almost anything I’ve read on the subject before. Why?

First, it doesn’t use the word “algorithm” at all. That is huge, because 95% of the time the use of the word “algorithmic” in the technology-and-society literature is stupid. People use “algorithm” when they really mean “software”. Now, they use “AI System” to mean “a company”. It’s ridiculous.

O’Neil makes it clear in this chapter that what she’s talking about are different kinds of models. Models can be in ones head (as in her plan for feeding her family) or in a computer, and both kinds of models can be racist. That’s a helpful, sane view. It’s been the consensus of computer scientists, cognitive scientists, and AI types for decades.

The problem with WMDs, as opposed to other, better models, is that the WMDS models are unhinged from reality. O’Neil’s complaint is not with use of models, but rather that models are being used without being properly trained using sound sampling on data and statistics. WMDs are not artificially intelligences; they are artificial stupidities.

In more technical terms, it seems like the problem with WMDs is not that they don’t properly trade off predictive accuracy with fairness, as some computer science literature would suggest is necessary. It’s that the systems have high error rates in the first place because the training and calibration systems are poorly designed. What’s worse, this avoidable error is disparately distributed, causing more harm to some groups than others.

This is a wonderful and eye-opening account of unfairness in the models used by automated decision-making systems (note the language). Why? Because it shows that there is a connection between statistical bias, the kind of bias that creates distortions in a quantitative predictive process, and social bias, the kind of bias people worry about politically, which consistently uses the term in both ways. If there is statistical bias that is weighing against some social group, then that’s definitely, 100% a form of bias.

Importantly, this kind of bias–statistical bias–is not something that every model must have. Only badly made models have it. It’s something that can be mitigated using scientific rigor and sound design. If we see the problem the way O’Neil sees it, then we can see clearly how better science, applied more rigorously, is also good for social justice.

As a scientist and technologist, it’s been terribly discouraging in the past years to be so consistently confronted with a false dichotomy between sound engineering and justice. At last, here’s a book that clearly outlines how the opposite is the case!

For fairness in machine learning, we need to consider the unfairness of racial categorization

Pre-prints of papers accepted to this coming 2019 Fairness, Accountability, and Transparency conference are floating around Twitter. From the looks of it, many of these papers add a wealth of historical and political context, which I feel is a big improvement.

A noteworthy paper, in this regard, is Hutchinson and Mitchell’s “50 Years of Test (Un)fairness: Lessons for Machine Learning”, which puts recent ‘fairness in machine learning’ work in the context of very analogous debates from the 60’s and 70’s that concerned the use of testing that could be biased due to cultural factors.

I like this paper a lot, in part because it is very thorough and in part because it tees up a line of argument that’s dear to me. Hutchinson and Mitchell raise the question of how to properly think about fairness in machine learning when the protected categories invoked by nondiscrimination law are themselves social constructs.

Some work on practically assessing fairness in ML has tackled the problem of using race as a construct. This echoes concerns in the testing literature that stem back to at least 1966: “one stumbles immediately over the scientific difficulty of establishing clear yardsticks by which people can be classified into convenient racial categories” [30]. Recent approaches have used Fitzpatrick skin type or unsupervised clustering to avoid racial categorizations [7, 55]. We note that the testing literature of the 1960s and 1970s frequently uses the phrase “cultural fairness” when referring to parity between blacks and whites.

They conclude that this is one of the areas where there can be a lot more useful work:

This short review of historical connections in fairness suggest several concrete steps forward for future research in ML fairness: Diving more deeply into the question of how subgroups are defined, suggested as early as 1966 [30], including questioning whether subgroups should be treated as discrete categories at all, and how intersectionality can be modeled. This might include, for example, how to quantify fairness along one dimension (e.g., age) conditioned on another dimension (e.g., skin tone), as recent work has begun to address [27, 39].

This is all very cool to read, because this is precisely the topic that Bruce Haynes and I address in our FAT* paper, “Racial categories in machine learning” (arXiv link). The problem we confront in this paper is that the racial categories we are used to using in the United States (White, Black, Asian) originate in the white supremacy that was enshrined into the Constitution when it was formed and perpetuated since then through the legal system (with some countervailing activity during the Civil Rights Movement, for example). This puts “fair machine learning” researchers in a bind: either they can use these categories, which have always been about perpetuating social inequality, or they can ignore the categories and reproduce the patterns of social inequality that prevail in fact because of the history of race.

In the paper, we propose a third option. First, rather than reify racial categories, we propose breaking race down into the kinds of personal features that get inscribed with racial meaning. Phenotype properties like skin type and ocular folds are one such set of features. Another set are events that indicate position in social class, such as being arrested or receiving welfare. Another set are facts about the national and geographic origin of ones ancestors. These facts about a person are clearly relevant to how racial distinctions are made, but are themselves more granular and multidimensional than race.

The next step is to detect race-like categories by looking at who is segregated from each other. We propose an unsupervised machine learning technique that works with the distribution of the phenotype, class, and ancestry features across spatial tracts (as in when considering where people physically live) or across a social network (as in when considering people’s professional networks, for example). Principal component analysis can identify what race-like dimensions capture the greatest amounts of spatial and social separation. We hypothesize that these dimensions will encode the ways racial categorization has shaped the social structure in tangible ways; these effects may include both politically recognized forms of discrimination as well as forms of discrimination that have not yet been surfaced. These dimensions can then be used to classify people in race-like ways as input to fairness interventions in machine learning.

A key part of our proposal is that race-like classification depends on the empirical distribution of persons in physical and social space, and so are not fixed. This operationalizes the way that race is socially and politically constructed without reifying the categories in terms that reproduce their white supremacist origins.

I’m quite stoked about this research, though obviously it raises a lot of serious challenges in terms of validation.

How to promote employees using machine learning without societal bias

Though it may at first read as being callous, a managerialist stance on inequality in statistical classification can help untangle some of the rhetoric around this tricky issue.

Consider the example that’s been in the news lately:

Suppose a company begins to use an algorithm to make decisions about which employees to promote. It uses a classifier trained on past data about who has been promoted. Because of societal bias, women are systematically under-promoted; this is reflected in the data set. The algorithm, naively trained on the historical data, reproduces the historical bias.

This example describes a bad situation. It is bad from a social justice perspective; by assumption, it would be better if men and women had equal opportunity in this work place.

It is also bad from a managerialist perspective. Why? Because if the point of using an algorithm were not to correct for societal biases introducing irrelevancies into the promotion decision, then it would not make managerial sense to change business practices over to using an algorithm. The whole point of using an algorithm is to improve on human decision-making. This is a poor match of an algorithm to a problem.

Unfortunately, what makes this example compelling is precisely what makes it a bad example of using an algorithm in this context. The only variables discussed in the example are the socially salient ones thick with political implications: gender, and promotion. What are more universal concerns than gender relations and socioeconomic status?!

But from a managerialist perspective, promotions should be issued based on a number of factors not mentioned in the example. What factors are these? That’s a great and difficult question. Promotions can reward hard work and loyalty. They can also be issued to those who demonstrate capacity for leadership, which can be a function of how well they get along with other members of the organization. There may be a number of features that predict these desirable qualities, most of which will have to do with working conditions within the company as opposed to qualities inherent in the employee (such as their past education, or their gender).

If one were to start to use machine learning intelligently to solve this problem, then one would go about solving it in a way entirely unlike the procedure in the problematic example. One would rather draw on soundly sourced domain expertise to develop a model of the relationship between relevant, work-related factors. For many of the key parts of the model, such as general relationships between personality type, leadership style, and cooperation with colleagues, one would look outside the organization for gold standard data that was sampled responsibly.

Once the organization has this model, then it can apply it to its own employees. For this to work, employees would need to provide significant detail about themselves, and the company would need to provide contextual information about the conditions under which employees work, as these may be confounding factors.

Part of the merit of building and fitting such a model would be that, because it is based on a lot of new and objective scientific considerations, it would produce novel results in recommending promotions. Again, if the algorithm merely reproduced past results, it would not be worth the investment in building the model.

When the algorithm is introduced, it ideally is used in a way that maintains traditional promotion processes in parallel so that the two kinds of results can be compared. Evaluation of the algorithm’s performance, relative to traditional methods, is a long, arduous process full of potential insights. Using the algorithm as an intervention at first allows the company to develop a causal understanding its impact. Insights from the evaluation can be factored back into the algorithm, improving the latter.

In all these cases, the company must keep its business goals firmly in mind. If they do this, then the rest of the logic of their method falls out of data science best practices, which are grounded in mathematical principles of statistics. While the political implications of poorly managed machine learning are troubling, effective management of machine learning which takes the precautions necessary to develop objectivity is ultimately a corrective to social bias. This is a case where sounds science and managerialist motives and social justice are aligned.

On achieving social equality

When evaluating a system, we have a choice of evaluating its internal functions–the inside view–or evaluating its effects situated in a larger context–the outside view.

Decision procedures (whether they are embodied by people or performed in concert with mechanical devices–I don’t think this distinction matters here) for sorting people are just such a system. If I understand correctly, the question of which principles animate antidiscrimination law hinge on this difference between the inside and outside view.

We can look at a decision-making process and evaluate whether as a procedure it achieves its goals of e.g. assigning credit scores without bias against certain groups. Even including processes of the gathering of evidence or data in such a system, it can in principle be bounded and evaluated by its ability to perform its goals. We do seem to care about the difference between procedural discrimination and procedural nondiscrimination. For example, an overtly racist policy that ignores truly talent and opportunity seems worse than a bureaucratic system that is indifferent to external inequality between groups that then gets reflected in decisions made according to other factors that are merely correlated with race.

The latter case has been criticized in the outside view. The criticism is captured by the phrasing that “algorithms can reproduce existing biases”. The supposedly neutral algorithm (which can, again, be either human or machine) is not neutral in its impact because in making its considerations of e.g. business interest are indifferent to the conditions outside it. The business is attracted to wealth and opportunity, which are held disproportionately by some part of the population, so the business is attracted to that population.

There is great wisdom in recognizing that institutions that are neutral in their inside view will often reproduce bias in the outside view. But it is incorrect to therefore conflate neutrality in the inside view with a biased inside view, even though their effects may be under some circumstances the same. When I say it is “incorrect”, I mean that they are in fact different because, for example, if the external conditions of procedurally neutral institution change, then it will reflect those new conditions. A procedurally biased institution will not reflect those new conditions in the same way.

Empirically it is very hard to tell when an institution is being procedurally neutral and indeed this is the crux of an enormous amount of political tension today. The first line of defense of an institution blamed of bias is to claim that their procedural neutrality is merely reflecting environmental conditions outside of its control. This is unconvincing for many politically active people. It seems to me that it is now much more common for institutions to avoid this problem by explicitly declaring their bias. Rather than try to accomplish the seemingly impossible task of defending their rigorous neutrality, it’s easier to declare where one stands on the issue of resource allocation globally and adjust ones procedure accordingly.

I don’t think this is a good thing.

One consequence of evaluating all institutions based on their global, “systemic” impact as opposed to their procedural neutrality is that it hollows out the political center. The evidence is in that politics has become more and more polarized. This is inevitable if politics becomes so explicitly about maintaining or reallocating resources as opposed to about building neutrally legitimate institutions. When one party in Congress considers a tax bill which seems designed mainly to enrich ones own constituencies at the expense of the other’s things have gotten out of hand. The idea of a unified idea of ‘good government’ has been all but abandoned.

An alternative is a commitment to procedural neutrality in the inside view of institutions, or at least some institutions. The fact that there are many different institutions that may have different policies is indeed quite relevant here. For while it is commonplace to say that a neutral institution will “reproduce existing biases”, “reproduction” is not a particularly helpful word here. Neither is “bias”. What we can say more precisely is that the operations of procedurally neutral institution will not change the distribution of resources even though they are unequal.

But if we do not hold all institutions accountable for correcting the inequality of society, isn’t that the same thing as approving of the status quo, which is so unequal? A thousand times no.

First, there’s the problem that many institutions are not, currently, procedurally neutral. Procedural neutrality is a higher standard than what many institutions are currently held to. Consider what is widely known about human beings and their implicit biases. One good argument for transferring decision-making authority to machine learning algorithms, even standard ones not augmented for ‘fairness’, is that they will not have the same implicit, inside, biases as the humans that currently make these decisions.

Second, there’s the fact that responsibility for correcting social inequality can be taken on by some institutions that are dedicated to this task while others are procedurally neutral. For example, one can consistently believe in the importance of a progressive social safety net combined with procedurally neutral credit reporting. Society is complex and perhaps rightly has many different functioning parts; not all the parts have to reflect socially progressive values for the arc of history to bend towards justice.

Third, there is reason to believe that even if all institutions were procedurally neutral, there would eventually be social equality. This has to do with the mathematically bulletproof but often ignored phenomenon of regression towards the mean. When values are sampled from a process at random, their average will approach the mean of the distribution as more values are accumulated. In terms of the allocation of resources in a population, there is some random variation in the way resources flow. When institutions are fair, inequality in resource allocation will settle into an unbiased distribution. While their may continue to be some apparent inequality due to disorganized heavy tail effects, these will not be biased, in a political sense.

Fourth, there is the problem of political backlash. Whenever political institutions are weak enough to be modified towards what is purported to be a ‘substantive’ or outside view neutrality, that will always be because some political coalition has attained enough power to swing the pendulum in their favor. The more explicit they are about doing this, the more it will mobilize the enemies of this coallition to try to swing the pendulum back the other way. The result is war by other means, the outcome of which will never be fair, because in war there are many who wind up dead or injured.

I am arguing for a centrist position on these matters, one that favors procedural neutrality in most institutions. This is not because I don’t care about substantive, “outside view” inequality. On the contrary, it’s because I believe that partisan bickering that explicitly undermines the inside neutrality of institutions undermines substantive equality. Partisan bickering over the scraps within narrow institutional frames is a distraction from, for example, the way the most wealthy avoid taxes while the middle class pays even more. There is a reason why political propaganda that induces partisan divisions is a weapon. Agreement about procedural neutrality is a core part of civic unity that allows for collective action against the very most abusively powerful.

References

Zachary C. Lipton, Alexandra Chouldechova, Julian McAuley. “Does mitigating ML’s disparate impact require disparate treatment?” 2017

Notes on fairness and nondiscrimination in machine learning

There has been a lot of work done lately on “fairness in machine learning” and related topics. It cannot be a coincidence that this work has paralleled a rise in political intolerance that is sensitized to issues of gender, race, citizenship, and so on. I more or less stand by my initial reaction to this line of work. But very recently I’ve done a deeper and more responsible dive into this literature and it’s proven to be insightful beyond the narrow problems which it purports to solve. These are some notes on the subject, ordered so as to get to the point.

The subject of whether and to what extent computer systems can enact morally objectionable bias goes back at least as far as Friedman and Nissenbaum’s 1996 article, in which they define “bias” as systematic unfairness. They mean this very generally, not specifically in a political sense (though inclusive of it). Twenty years later, Kleinberg et al. (2016) prove that there are multiple, competing notions of fairness in machine classification which generally cannot be satisfied all at once; they must be traded off against each other. In particular, a classifier that uses all available information to optimize accuracy–one that achieves what these authors call calibration–cannot also have equal false positive and false negative rates across population groups (read: race, sex), properties that Hardt et al. (2016) call “equal opportunity”. This is no doubt inspired by a now very famous ProPublica article asserting that a particular kind of commercial recidivism prediction software was “biased against blacks” because it had a higher false positive rate for black suspects than white offenders. Because bail and parole rates are set according to predicted recidivism, this led to cases where a non-recidivist was denied bail because they were black, which sounds unfair to a lot of people, including myself.

While I understand that there is a lot of high quality and well-intentioned research on this subject, I haven’t found anybody who could tell me why the solution to this problem was to stop using predicted recidivism to set bail, as opposed to futzing around with a recidivism prediction algorithm which seems to have been doing its job (Dieterich et al., 2016). Recidivism rates are actually correlated with race (Hartney and Vuong, 2009). This is probably because of centuries of systematic racism. If you are serious about remediating historical inequality, the least you could do is cut black people some slack on bail.

This gets to what for me is the most baffling aspect of this whole research agenda, one that I didn’t have the words for before reading Barocas and Selbst (2016). A point well-made by them is that the interpretation anti-discrimination law, which motivates a lot of this research, is fraught with tensions that complicate its application to data mining.

“Two competing principles have always undergirded anti-discrimination law: nondiscrimination and antisubordination. Nondiscrimination is the narrower of the two, holding that the responsibility of the law is to eliminate the unfairness individuals experience a the hands of decisionmakers’ choices due to membership in certain protected classes. Antisubordination theory, in contrast, holds that the goal of antidiscrimination law is, or at least should be, to eliminate status-based inequality due to membership in those classes, not as a matter of procedure, but substance.” (Barocas and Selbst, 2016)

More specifically, these two principles motivate different interpretations of the two pillars of anti-discrimination law, disparate treatment and disparate impact. I draw on Barocas and Selbst for my understanding of each:

A judgment of disparate treatment requires either a formal disparate treatment (across protected groups) of similarly situated people, or an intent to discriminate. Since in a large data mining application protected group membership will be proxied by many other factors, it’s not clear if the ‘formal’ requirement makes much sense here. And since machine learning applications only very rarely have racist intent, that option seems challengeable as well. While there are interpretations of these criteria that are tougher on decision-makers (i.e. unconscious intents), these seem to be motivated by antisubordination rather than the weaker nondiscrimination principle.

A judgment of disparate impact is perhaps more straightforward, but it can be mitigated in cases of “business necessity”, which (to get to the point) is vague enough to plausibly include optimization in a technical sense. Once again, there is nothing to see here from a nondiscrimination standpoint, though a nonsubordinationist would rather that these decision-makers have to take correcting for historical inequality into account.

I infer from their writing that Barocas and Selbst believe that nonsubordination is an important principle for nondiscrimination. In any case, they maintain that making the case for applying nondiscrimination laws to data mining effectively requires a commitment to “substantive remediation”. This is insightful!

Just to put my cards on the table: as much as I may like the idea of substantive remediation in principle, I personally don’t think that every application of nondiscrimination law needs to be animated by it. For many institutions, narrow nondiscrimination seems to be adequate if not preferable. I’d prefer remediation to occur through other specific policies, such as more public investment in schools in low-income districts. Perhaps for this reason, I’m not crazy about “fairness in machine learning” as a general technical practice. It seems to me to be trying to solve social problems with a technical fix, which despite being quite technical myself I don’t always see as a good idea. It seems like in most cases you could have a machine learning mechanism based on normal statistical principles (the learning step) and then use a decision procedure separately that achieves your political ends.

I wish that this research community (and here I mean more the qualitative research community surrounding it more than the technical community, which tends to define its terms carefully) would be more careful about the ways it talks about “bias”, because often it seems to encourage a conflation between statistical or technical senses of bias and political senses. The latter carry so much political baggage that it can be intimidating to try to wade in and untangle the two senses. And it’s important to do this untangling, because while bad statistical bias can lead to political bias, it can, depending on the circumstances, lead to either “good” or “bad” political bias. But it’s important, from the sake of numeracy (mathematical literacy) to understand that even if a statistically bad process has a politically “good” outcome, that is still, statistically speaking, bad.

My sense is that there are interpretations of nondiscrimination law that make it illegal to make certain judgments taking into account certain facts about sensitive properties like race and sex. There are also theorems showing that if you don’t take into account those sensitive properties, you are going to discriminate against them by accident because those sensitive variables are correlated with anything else you would use to judge people. As a general principle, while being ignorant may sometimes make things better when you are extremely lucky, in general it makes things worse! This should be a surprise to nobody.

References

Barocas, Solon, and Andrew D. Selbst. “Big data’s disparate impact.” (2016).

Dieterich, William, Christina Mendoza, and Tim Brennan. “COMPAS risk scales: Demonstrating accuracy equity and predictive parity.” Northpoint Inc (2016).

Friedman, Batya, and Helen Nissenbaum. “Bias in computer systems.” ACM Transactions on Information Systems (TOIS) 14.3 (1996): 330-347.

Hardt, Moritz, Eric Price, and Nati Srebro. “Equality of opportunity in supervised learning.” Advances in Neural Information Processing Systems. 2016.

Hartney, Christopher, and Linh Vuong. “Created equal: Racial and ethnic disparities in the US criminal justice system.” (2009).

Kleinberg, Jon, Sendhil Mullainathan, and Manish Raghavan. “Inherent trade-offs in the fair determination of risk scores.” arXiv preprint arXiv:1609.05807 (2016).