Reflections on Fair Machine Learning by Christian Fricke, Lead Back-end Developer at Saidot

Who I Am

Hello. My name is Christian Fricke. I am a computer scientist and software developer, primarily interested in furthering the discussion of ethics in machine learning and data mining, free and open source software, programming language theory and philosophy. In short, I enjoy to be endlessly puzzled by the incompatibility of formalised logic and human behaviour.

Whilst studying Mediensysteme (Computer Science) at the Bauhaus-Universität Weimar, I became fascinated with the power of scale that machine learning applications offer for natural language processing and computer vision. After finishing my studies, to hone my software development and agile business skills, I I joined Unata Inc., a small startup in Toronto, which was set on disrupting the then leading market monopoly in the digital grocery space, by utilising machine learning for a 1-to-1 personalised platform for grocers.

Not long after, the issue of fairness in automated decision-making applications became a much-discussed topic, and an inevitable challenge to be tackled, which led me to move to Finland to pursue a Masters in Machine Learning and Data Mining at Aalto University, Helsinki. One year later, I joined Futurice as a software and data science consultant. In this role, I promoted governmental and public good contracts, and worked together with Minna Mustakallio (Head of Product at Saidot) on increasing awareness for AI ethics. Since the topic of my ongoing thesis work, fair machine learning, fairness-aware data mining and ethics in AI, seemed to naturally align with Saidot’s goal, I joined the company in spring 2019.

The State of Fair Machine Learning

This area of Machine Learning (ML) is still fairly young, and complex by its nature, as it influences and is influenced by areas such as legislation and sociology. Many currently deployed AI technologies and applications run unchecked, and are in dire need of regulation and contextual verification. Dwork et al.'s contributions, among others, have increased the awareness and created the field of Fair ML almost a decade ago. Not much later, the FAT/ML forum and ACM Conference on Fairness were also founded.

Currently, there exist many different definitions for fairness in machine learning. However, the problem is that this definition varies with each one. According to Pedreshi et al,the main emphasis here should be on discrimination; this referring to the unfair or unequal treatment of people based on their membership in a category or a minority, without regard to individual merit. So, to tackle this issue, we have two choices: we can either detect unfairness by identifying unfair treatment in data sets and blackbox models, or prevent unfairness by learning a statistical model from potentially unfair data sets via pre-/in-/post-processing, or within certain tasks, such as regression.

When it comes to the causes of unfairness, we have to look at the biases embedded in our data sets and applications. We deal with:

1. annotation bias, when protected attributes are encoded in observable variables.
2. sample bias, inherent to human created data, where the majority of data sets include only target specific records, ignoring otherwise affected subgroups.
3. inductive bias, the basis of machine learning, caused by assumptions made by the algorithm in order to learn generalisations.

There are many other important types of bias, including errors that concentrate in a specific class, incorrectly labeled data and oversimplified models, which all lend themselves to the various definitions of fairness. Kamishima offers the following definition of formal fairness:

“the desired condition defined by a formal relation between sensitive feature, target variable, and other variables in a model”.

He further enumerates the different types as direct, indirect, group, individual and symmetric fairness. The literature provides more than 40 definitions of fairness measures in total, with the most widely used ones already proven incompatible with one another, as shown by Corbett-Davies and Goel. I will elaborate on all of these concepts in future posts.

Right now, fair machine learning takes center stage not only in politics, private and public organisations, and AI research, but also in the lives of citizens, who are already being affected by automated decision making on a daily basis. As technologists we have to take a closer look at the problem of model accuracy over fairness and what it actually means to treat everyone equally, or equitably. The utility of scale that machine learning currently offers through generalisation affects minorities directly. Therefore, as scientists we cannot simply quantify the well-being of these individuals based on formal, rigorous mathematical definitions. Instead, we need to actively engage with lawmakers and social scientists to not merely accommodate the majority and the mainstream, but also the marginalised and the underprivileged.

Back to the Drawing Board

At the end of May, I will participate in the International Summer School on Bias and Discrimination in AI (Montreal, Canada), where, as part of the audience, I hope to experience thought-provoking lectures and engage in active discourse on the contextual nature of ethics and fairness. I am excited to learn more about the current measures of formal fairness and on the ways to handle the various forms of bias beyond doctrines such as disparate impact and disparate treatment. I plan to share my findings and reflections from the conference with the Big Data research group at Aalto University and others to help facilitate Saidot’s role as a leading example.

If you'd like to have Christian as your colleague, check out our open positions in the development team here.