It's not that O'Neil is against big data per se. She points out examples where it can be useful and effective - but this requires the systems to be transparent and to be capable of learning from their mistakes. In the examples we discover, from systems that rate school teachers to those that decide whether or not to issue a payday loan, the system is opaque, secretive and based on a set of rules that aren't tested against reality and regularly updated to produce a fair outcome.
The teacher grading system is probably the most dramatically inaccurate example, where the system is trying to measure how well a teacher has performed, based on data that only has a very vague link to actual outcomes - so, for instance, O'Neil tells of a teacher who scored 6% one year and 96% the next year for doing the same job. The factors being measured are almost entirely outside the teacher's control with no linkage to performance and the interpretation of the data is simply garbage.
Other systems, such as those used to rank universities, are ruthlessly gamed by the participants, making them far more about how good an organisation is at coming up with the right answers to metrics than it is to the quality of that organisation. And all of us will come across targeted advertising and social media messages/search results prioritised according to secret algorithms which we know nothing about and that attempt to control our behaviour.
For O'Neil, the worst aspects of big data misuse are where a system - perhaps with the best intentions - ends up penalising people for being poor of being from certain ethnic backgrounds. This is often a result of an indirect piece of data - for instance the place they live might have implications on their financial state or ethnicity. She vividly portrays the way that systems dealing with everything from police presence in an area to fixing insurance premiums can produce a downward spiral of negative feedback.
Although the book is often very effective, it is heavily US-oriented, which is a shame when many of these issues are as significant, say, in Europe, as they are in the US. There is probably also not enough nuance in the author's binary good/bad opinion of systems. For example, she tells us that someone shouldn't be penalised by having to pay more for insurance because they live in a high risk neighbourhood - but doesn't think about the contrary aspect that if insurance companies don't do this, those of us who live in low risk neighbourhoods are being penalised by paying much higher premiums than we need to in order to cover our insurance.
O'Neil makes a simplistic linkage between high risk = poor, low risk = rich - yet those of us, for instance, who live in the country are often in quite poor areas that are nonetheless low risk. For O'Neil, fairness means everyone pays the same. But is that truly fair? Here in Europe, we've had car insurance for young female drivers doubled in cost to make it the same as young males - even though the young males are far more likely to have accidents. This is fair by O'Neil's standards, because it doesn't discriminate on gender, but is not fair in the real world away from labels.
O'Neil makes a simplistic linkage between high risk = poor, low risk = rich - yet those of us, for instance, who live in the country are often in quite poor areas that are nonetheless low risk. For O'Neil, fairness means everyone pays the same. But is that truly fair? Here in Europe, we've had car insurance for young female drivers doubled in cost to make it the same as young males - even though the young males are far more likely to have accidents. This is fair by O'Neil's standards, because it doesn't discriminate on gender, but is not fair in the real world away from labels.
There's a lot here that we should be picking up on, and even if you don't agree with all of O'Neil's assessments, it certainly makes you think about the rights and wrongs of decisions based on automated assessment of indirect data.
Paperback: 
Kindle:
Using these links earns us commission at no cost to you
Using these links earns us commission at no cost to you
Review by Brian Clegg
This book is basically a discussion of the negative effects of 'unfair' models when applied to people, with copious examples given. The author has been working as a data scientist, a role that I split into two types - those who can explain their statistical models and those who can't explain their machine learning (ML) models. Note that sometimes a statistical model can be derived from an ML model. Most of the models are described as black boxes - presumably ML models.
ReplyDeleteThe author confesses to being an SJW although that doesn't mean that at times her ire isn't justified - thankfully the book doesn't read as a polemic. She's an American and thinks that it was a good idea that the British NHS stopped rejecting doctors for poor English language skills. Unfortunately, according to the local news, this has led to patients dying. She worked on Wall Street during the crash of 2007, which she is rightly critical of. Interestingly she is a fan of Barack Obama even though the then Senator didn't support the 2005 Bill that would have stopped Fannie Mae and Freddie Mac from making those subprime loans.
There is very little technical analysis beyond discussion of the inputs and weightings. A/B testing is presented early on in a very negative light and near the end mentioned again in a better light. In the conclusion it is suggested that the amount of data could be reduced to sacrifice accuracy for fairness with no mention that it could instead increase the quality of predictions by eliminating over-fitting. So not really my sort of book, I prefer less data and more technical details.