The Importance of Diversity in AI & Machine Learning

Published: May 24, 2023

Michelle McGlone — Lead Data Scientist — D3 Claims Casualty

Hi, my name is Michelle McGlone and I am a Lead Data Scientist in the D3 department within Allstate Northern Ireland. I've been with Allstate since 2015 and currently take the lead on our data science workstreams in the Claims Casualty space. Our team has numerous high-value projects across Claims Casualty that act to deliver insights to claim handlers, which then result in the more efficient opening and management of injury coverages, reduced rates of attorney representation and litigation, and overall reduced loss costs and expenses.

Why is diversity so crucial to the development of Artificial Intelligence and Machine Learning

Behind every data point is a person, and the analytical solutions we build are ultimately used to impact a wide array of people ranging from our own broad customer base to prospective customers and third-party claimants. We need to ensure that the data going into our machine learning models is a diverse as the population that our solutions are applied to. A predictive model is only as good as the data it is trained on, and the more diverse the data, the better the performance of our models will be.

It is also important then that we consider any potential historical trends or bias in the data we collect. Building a model to predict future events could be biased towards repeating those same trends. Use cases and consequences of this must be heavily considered as to how historical trends should be handled within the data, and then in how a model prediction would be used in the real world — will it be automating action, or providing insight or guidance? This is highly important to consider while gathering data and building a modelling dataset.

Finally, building AI & ML products is only one piece of the puzzle — the means of implementation and how model insights reach their end goal also requires diverse teams. It is beneficial for all to encourage building teams with diverse experiences and perspectives to allow for out-of-the-box thinking and well-rounded processes to be put in place when setting up AI & ML pipelines. Ultimately, this is the piece of the puzzle that will drive results and change behaviours and actions, so it's crucial to do what we can to widen our gaze for our products through to the finish line.

How can we ensure fairness and diversity in our handling of data and the development of Al & ML technology?

Within Allstate, we have high levels of governance applied across the data & analytics solutions that we provide. These processes ensure our data and modelling pipelines are adhering to high standards of data collection, retention, development and suitability for use case. Throughout the build and processing of a dataset, data should be tested and cross-examined to validate that distributions are as expected and are not skewed or biased in any unexpected way that could subvert the goal the model aims to achieve.

Validation of the data continues prior to and following ingestion by a model. Forms of shadow testing and small-scale tests are extremely useful to simulate how a model will act in a production setting before continuing to wide-scale deployment. Following deployment, close monitoring of data distributions and model outputs continues to be important to identify bugs or risks that could cause adverse impact. With regular tracking and monitoring, we can quickly act to make changes where needed and continue to ensure our models are reaching their full potential.

The area of explainable AI is also helping to achieve transparency of machine learning models. While this is available at different levels depending on the model structures being used, interpretability of ML can give detailed insight into the inner workings of an ML process and how it leads to the resulting output. This can help data scientists validate a model is working as intended, as well as helping consumers of the model output understand and increase trust in the models.

A diverse Al workforce made up of different races, genders, ethnicities and ages decreases the likelihood of racial, gender, ethnic and age discrimination by artificially intelligent systems. How can we encourage and develop a diverse workforce within Allstate?

This is something that is important in any field of work, but especially so when building and implementing machine learning solutions that will ultimately impact people on the other end. It is important to build diverse teams with unique perspectives so that various angles are fully thought through before a product reaches its final stage, so that we ensure all parts of the puzzle are built fairly and are relevant to the use case from beginning to end.

There are many ways we can achieve and encourage diverse teams — the first being that we need to ensure we are casting a wide diverse net during out recruitment processes. There was once a time when all data scientists came straight from a mathematics/physics background, due to the statistical education required. The industry has grown a lot in recent years and there are now many different academic and industrial routes that can lead someone to a career in data science — AI. Companies must therefore ensure that recruitment processes reflect that. This includes making sure role requirements and interview processes are not hyper-specific to certain intake routes, and that we ensure our university and industry outreach caters to a wide array of potential employees who come in with the right skillset and diverse experiences and perspectives. The world of remote working also benefits us in this area as it allows recruitment to expand outside of our local markets and benefit from a wide array of experience across the sector.

Getting people in the door is one thing, but it is also very important to then be able to foster a diverse workforce. The existence of programs such as employee resource groups and mentoring/coaching opportunities for underrepresented sectors of our workforce can greatly benefit the empowerment of our employees and the encouragement for us all to live into and nourish a diverse working environment.