Algorithms have never been more publicly debated than in the last couple of weeks. As data science experts, it didn’t come as a surprise to us that an algorithm was used by the government to aid in the grading of students’ A-level grades in this extraordinary year. The use of data science to set (or help set) grades is, in principle, a sound idea. What did surprise us was the way in which the algorithm was developed and communicated to students and the public. This was likely a key factor in why so many people across the country were dissatisfied with the results it generated.
It all comes back to a question of ethics and empathy. Data science is most likely one of the last subjects where empathy would seem important. However, it’s this misinterpretation that often leads to the breakdown of data science projects – the model works, but does it consider the problem from the right perspective? Is it answering the right question? After all, empathy isn’t just about compassion or sympathy, it’s the ability to see a situation from someone else’s frame of reference.
So, in this instance, the A-level algorithm may have worked technically (without access to the detailed workings or results of the model we can’t say), insofar that it produced results accurate at the overall population level in comparison to previous years’ results. However, was the question of empathy taken into account, and therefore was accuracy the right thing to be aiming for?
We asked our Deputy Director of the data science consulting team, Dave Gardner, for his thoughts on the use of data science in assessing school performance:
“Without being privy to either the detailed results the algorithm generated, or the process by which is was built, it is impossible to say exactly what went wrong. We can however examine some of the decisions that we do know were made from the point of view of empathic and ethical data science.
“If part of your objective is to produce a set of results as similar as possible to that of previous years then the idea of incorporating past school performance to your algorithm does have some merit. Speaking from a purely technical perspective it’s a reasonable path to take – some schools may be optimistic, pessimistic, or interpret guidelines in different ways.
“The core issue however wasn’t with the algorithm’s accuracy at the population level, but with its (perceived) fairness at an individual level. Whether it is fair to give students a set of grades calculated by an algorithm to be the most likely outcome of a normal exam process, based partly on historic performance at the school level, isn’t a question data science can answer. However, it absolutely should have been a question it asked.
“Similarly the decision not to incorporate teacher’s predicted grades was clearly contentious but the thinking behind this decision was not well explained, and appeared to have been made without consultation with the right stakeholders.
“Had the explicit focus been on generating the fairest possible set of results for each individual (for the right definition of fairest) then I suspect we would have ended up with a very different, and much better received, algorithm.”
It’s no secret that there has been some amazing work done through the power of data science and analytics – but the lesson to be learnt here is the importance of understanding that there is a difference between accurate and fair data science, and at the end of the day, empathy is key.
Author: Dave Gardner, Deputy Director, Mango Solutions