The Simility Blog

Overcoming The Challenge Of Bias And Discrimination In Machine Learning
Jonathan Doering
Jonathan Doering
January 30, 2020

After Phillip Alston, UN rapporteur on poverty and human rights, presented his report to the UN a few weeks ago, the machine learning (ML) community sat up and took notice. In his report, Alston pointed out how new technologies revolutionize the interaction between governments and society and proclaimed later, that artificial intelligence (AI) needs to follow more human principles to earn the predicate “intelligent”. Many in the industry agreed with him and expressed concerns about how big data and machine learning can rather fuel a problem than solve it.

The ethics debate in AI is not a new one, though in the eyes of many AI researchers a very much underrated one. In times where machine learning is shaping daily life faster than ever before, the issue of biased machines making decisions becomes more and more problematic. Not only for consumers, but also for companies.

Some memorable examples demonstrate this issue that constantly happens to companies aiming to leveraging machine learning in their workflows. I remember one case where ML models used by healthcare providers to assign care to millions of people underestimated how urgently black people needed care and as a consequence, put white patients first. Another mishap involved a large marketplace that abandoned an in-house developed AI tool used for hiring, after it was found to rate male applicants much higher than female applicants.

Having an impact on people’s lives with our work within the Simility team, we are very aware of the problem of machine learning model bias. Since starting the company 5 years ago with machine learning always being a core competence, Simility implemented several principles into the platform to cope with the challenge of bias and discrimination in machine learning.

The First Challenge: Bias

One of the biggest challenges towards better and less biased machine learning solutions, is the problem of data bias and human lensing. Since our experiences shape the way we see and interact with our environment, what we have learned determines pretty much our work and its outcome. This touches a variety of areas in the data science workflow. What features are used, how complex the model becomes, and how this outcome is utilized for decision making.

A fundamental Simility vision was always around this challenge of data teams, and hence we have built in checks and balances in our platform to enable better building and deployment.

For example, the auditing function utilizes the four-eye principle prior to model production and consequently helps to reduce individual bias. Running multiple models in parallel within Simility, some of them in stealth mode, helps data scientists to monitor and understand model behavior in detail before utilized for decisioning.

The Second Challenge: Unintelligibility

Another key for better and more ethical machine learning lies in the selection and understanding of the input data itself. Simility’s engine generates thousands of very well defined, anonymized though clear to interpret features, in real time. These features are the foundations for the machine learning models, and as such gives data scientists the possibility to well understand what the model is using for decisioning. Usually, machine learning models suffer from a trade-off between complexity and interpretability. With easy to understand and comprehend input features, even decision making with more complex algorithms such as Gradient Boosted machines becomes interpretable.

And even in cases where better to interpret algorithms, such as general linearized models or logistic regression might be preferable, the presence of powerful features becomes even more important.

The Third Challenge: Discrimination

Discrimination from ML models is a very serious problem and becomes more and more important and visible where machine learning is increasingly used to make decisions across social domains. There is a paradigm shift in the usage of these algorithms, but people have been slower in understanding the impact of these decisions. As the general population is catching up with the rapidly changing aspect of algorithmic decision making and these decisions impact the day-to-day lives of people, we need to be proactive in making our decision strategies free from discrimination. Just taking the examples discussed earlier, a model built to improve healthcare but one that discriminates against an ethnical group, obviously creates more problems than it solves.

In banking and e-commerce, two businesses where Simility serves many customers, this is very viable as well. Banks need to make sure that customers are treated equally when it comes to, for example, gender in loan applications. And, e-commerce companies can’t deny orders only by geographical background.

There are different concepts to tackle discrimination problems in machine learning models. One of those concepts is the “equal opportunity threshold” introduced by Moritz Hardt et al. The equal opportunity threshold basically aims to set the decision criteria to a level, where for each group the same fraction profits from the decision. This concept has the advantage, that it can be introduced without changing the model itself – just by adjusting the decision threshold. With Simility’s AutoDecision Tabs, setting different thresholds for different scenarios becomes easy and it allows for flexible decision making on a variety of different criteria with easy trackability.

The Fourth Challenge: Black-Box Models

Though a deep understanding of the model decisioning process during the development phase is a very sound pillar towards more ethical machine learning, it is still a necessity to constantly monitor and understand model reasoning on a case-by-case basis. As this is one of the most active research areas, over the past years several novel techniques have been developed which try to understand why a certain model decides for a particular outcome for a given sample.

In production, explainability methods such as LIME, Shapley or RL-LIM, which have very solid theoretical underpinnings, help to understand what is happening on a sample-by-sample basis. Model explainer helps to put trust into the machine’s decision process and allows for better understanding, leading to better knowledge gained from the data. Meaningful explanations enable fraud operation teams and data scientists equally, to identify new emerging fraud patterns faster which helps them to develop powerful counter attacks quicker and eventually positively optimize the overall business and operational KPIs.

Simility’s prediction engine leverages exactly such methods, and outputs an interpretability plot for each event.

In addition, more and more governmental regulations are taking place and enforcing companies to make machine decisioning processes transparent for and explainable to customers upon request. Simility’s explainability plots deliver valuable insights here and can help in communication with regulatory agencies.

Developing trustable and unbiased machine learning models is not an easy task. Many proposals are floating around, from certification for unbiased data sets to a dedicated oath for data scientists. To date, there is not one single way that leads the path. Machine learning models are still statistical models, which learn numerical patterns from input data to produce predictable outcomes. Until we develop models which unveil their very own decisioning process in a clear, humanly understandable way, we have to rely on proxy techniques and intuition. Having a technology that creates transparent features and offers model explainability in real time, enables us at Simility to develop first-class models we understand. Remember, don’t deploy what you have not understood.

To learn more about Simility’s industry recognized machine learning platform, schedule a demo now.