How do you share machine learning insights with non-technical audiences?

One of the most helpful things to know about your audience is what their incentives are. In many cases, this translates to how they get paid, or how they get bonuses and promotions. In some other cases, their incentives are not necessarily monetary: save the whales, get elected, increase their audience. In any case, knowing what motivates them is very helpful to tailor your communication.

Focus on the business. That’s it. The problem, the outcomes. And sprinkle the aha somewhere in between. We often think our technical competence is a gift that we need to awe an audience with. In reality, most amazing solutions are amazingly easy to relate to, as long as we focus on why the model was built in the first place

Understanding your audience means going beyond demographics. It's about empathising with their concerns and linking machine learning outcomes to their real-world needs. Simplify technical concepts, draw parallels to everyday experiences, and create relatable narratives. This paves the way for communication that's not just information but a conversation.

In my experience, avoid technical complexity over understanding. What might be obvious to you might not be at your audience reach. Maximize your audience objective. Communication is also part of your job as a modeler.

One effective way to know your audience is to utilize existing survey data from your target group, or perform an on-site assessment while giving a talk. This way you can provide examples on how the outcomes can be used in case studies the audience can relate to.

In my experience, it's usually better to show the solution first and then explain how you got to it. Most non-technical audiences don't care about the deeper details of how you got to a result, but showing the results first at least help their anchor their thinking, especially if the result is surprising.

It's vital to make a clear distinction with non-technical partners about which problems can be effectively addressed by developing machine learning models and which cannot. This consideration is the primary step before delving into more specific details. I've witnessed scenarios where teams squandered hundreds of hours due to a lack of this careful consideration.

Verbal intuitive explanations of machine learning interpretability results like SHAP values is crucial. You would need to set the tone for complexity and comprehensiveness according to the audiance and use case. In addition, it is essential to talk about limitations of the model, like lack of extrapolation if a tree based algorithm is chosen etc

It is also important to explain the limitations of approaches other than machine learning to justify the use of such a laborious approach.

This should depend on the knowledge that your target audience has. If the target audience is non-technical people, it was better to show clear motivation on why the problem at had is interesting, and then present the solution.

People need information not to look at it, but to make decisions based on it. When visualizing information, I follow a simple rule: differentiate the colors, shapes, and position of shapes in space on a slide based on a pool of decisions that the person studying your slide makes. For example, you can tint different paths based on their degree of risk from red to green, or mark expenses with circles of different sizes.

In my experience, each illustration should be screaming one take-home message and not require more than a minute to decipher.

As you interact with many clients from distinct domains, you observe that people have an appetite for colorful representations. For instance, a full spectrum palette heatmaps somehow fascinate and engage people though they are actually not that interpretable so much compared to a binned graph with a monochromatic palette.

Every new problem requires new set of graphs. There are no two problems alike, same goes for graphs. Aim for the simplest graphs, make text and numbers readable, make sure to use the right palette of colors that are inclusive and publication ready (generally in shades of gray). Avoid common mistakes and make sure graphs are comparable, and at the same scale. For example, if you are mapping your numbers to height, make sure that twice the height means twice the quantity.

We create our AI like a student preparing for a test. We teach it (train it), and then it takes several practice exams (validation) to ensure it's ready. When it makes mistakes, we figure out why (error analysis), like a tutor analyzing areas to improve. We examine which lessons affect its answers the most (explainability). We compare it to other students (model comparisons) to check its performance. All steps are documented (reproducibility) like a detailed study plan. We continually check its performance (continuous monitoring), verify that all subjects are equally treated(fairness), check the results under stress (robustness), and use analogies to convoy your message and it is an iterative process, like learning, requiring perseverance.

For instance, if we've developed a predictive model for customer churn in a telecom company, instead of diving into complex algorithms and mathematical formulas, we can have a conversation like, "Imagine our model as a system that helps us foresee which customers might leave us. It's like having a 'customer retention radar' that spots potential issues early on. This way, we can take proactive steps to keep our valued customers happy and retain for the long run." By using relatable analogies and real-life scenarios, we bridge the gap between data science jargon and its practical implications, making it easier for non-technical audiences to grasp the value and insights our models provide.

As we all already witnessed before, stories have been used throughout human history to connect and engage with others. By creating a narrative around your insights and findings, you can make the information more relatable and emotionally compelling. This human connection can foster a deeper understanding and empathy from your audience. Stories are inherently engaging and captivating, which is something you absolutely need when explaining more difficult concepts.

In my presentations, I always begin with a simple model to deliver key insights. As an economist, I refer to this model as a "dumbed-down" model, which may not capture the realism in the data but still carries useful insights. If you cannot explain how your model works, just being buried with numerous features using black-box algorithms, I don't know how you can persuade people to buy into your story. Can you convince financial regulators inspecting your consumer default model without understanding it yourself?

Recommendations must inspire confidence and offer tangible paths forward. Align them with your audience's goals, and emphasize how they lead to actionable solutions. Providing a roadmap, offering support, or involving the audience in the next steps can build trust and facilitate collaboration, turning insights into shared endeavours.

Translating machine learning insights into actionable recomendations is vital for delivering business outcomes. Example: If your customer segmentation model identifies a 'High-Value' group, advise launching targeted email campaigns that offer early access, explaining that this segment yields a higher ROI. Complement these actions with a feedback loop, data audits, and cross-channel extensions. By making these recommendations clear and explainable, you equip your stakeholders to make data-driven decisions that align with business objectives.

Always be open to feedback and ready to adapt. The process of communicating machine learning insights should be iterative and dynamic, reflecting the evolving needs and understanding of your audience. Engage in continuous dialogue, and view your insights not as a final product but a starting point for collaboration, learning, and growth.

Be approachable and actively listen to what your stakeholders have to say, while complementing their input with your expertise in data-driven decision-making. Example: For lifetime value (LTV) predictions, if a certain customer cohort is projected to spend 20% more within a year, recommend diverting more budget to retention tactics, like a tiered loyalty program, while underscoring the long-term ROI gains.

In my opinion, it’s best to be transparent about limitations/traps with your solution. It allows folks to know when/how to use your solution best, and establishes credibility and trust in the partnership.

Be open-minded to the people who can present and explain their domain knowledge in the general public. Such individuals teach us the art of communication, transforming specialized knowledge into accessible insights. I have really learned a lot from them Usually, you can find these people on TED lecture, but my favorite group of people are professors who would like to share and explain their research in a more relatable way with the audience.

one of the ways to share machine learning insights with non-technical audiences is by providing the benefits of machine learning based on their field of work.