Algorithmic Fairness, Bias, and Justice

In The Edinburgh Experience
5 months ago
2 Min read

I submitted the final paper for the course (and the semester) for my ‘Algorithmic Fairness, Bias, and Justice’ course. The final paper was titled Ethical Algorithms in Education: A Critical Analysis of Ofqual’s A-Level Grading Solution. Below is a snippet from the introduction section of the paper for a bit of context:

In 2020, considering the unprecedented COVID-19 public health emergency, the former Education Secretary Gavin Williamson announced that the summer exam series would be cancelled to help fight the spread of coronavirus, preventing students from sitting their A-level exams (Department for Education, 2020). Without actual exams, the Office of Qualifications and Examinations Regulation (Ofqual) estimated A-level grades using an algorithm. Ofqual’s algorithm has been widely reported as biased and unfair in its treatment of disadvantaged students and schools (Eccles & Gallardo, 2020; Kolkman, 2020). This report critically examines two specific issues originating from the Problem Specification phase of the algorithm’s design process (Figure 1). 

Overall, I found the course very useful. The course focuses on identifying and dealing with bias and fairness practically, including writing reports and designing interventions when appropriate. A few highlights from the course and required reading include:

  1. As with any powerful technology, machine learning raises questions about which of its potential uses society should encourage and discourage.
  2. Statistical patterns that apply to the majority might be invalid within a minority group.
  3. Achieving fairness might be computationally expensive if it forces us to look for more complex decision rules. Here’s an example:

    There might be a simple linear function that classifies the majority group correctly and a (different) simple linear function that classifies the minority group correctly, but learning a (non-linear) combination of two linear classifiers is computationally a harder problem. There are excellent algorithms available for learning linear classifiers (e.g., SVM), but no efficient algorithm that I know of for learning an arbitrary combination of two linear classifiers.

Read more