How do you explain your data science models?

Let's say I am looking for cereals on a retail eCommerce site. The retail search predicts which cereals are more likely to be purchased by customers. The underlying data science model learns the relationship between factors that influence cereal purchases, such as: customer's age, gender, zip code, and past purchase history. For e.g., here the model might learn, customers who are younger are more likely to buy sugary cereals, while older age groups may go with healthier cereals. When explaining models, it is important to keep it clear and use real-world examples to help understand the audience how the model works. Effective use of ML models in businesses can provide better customer experience, improve decision-making, and drive sales.

In my experience teaching AI to over 150k learners via LinkedIn Learning, it's best to use props to illustrate probabilities in AI systems. I often use a raffle to illustrate how a generative model might come up with the next token.

Certainly. I concur. Simplifying your model’s results for business comprehension is vital. Imagine explaining it to a child, and use visuals. Instead of stating an R-squared of 0.77, you might convey the model’s accuracy as 80% and depict it using a straightforward barometer graphic. In a time series forecasting model, instead of talking about ARIMA or exponential smoothing, you might say, “Our model predicts that sales will increase by about 10% next month based on past trends.”

Something I try to remember in this space is that if I can't explain a technique in simple language, I probably don't understand it fully. You can be the smartest person in the world, but unless you can translate that to your audience, your intelligence is worthless.

Absolutely! Employing plain language can have a significant impact. I strive to utilize different real-life situations to elucidate the intricacies of the technology.

Something interesting about the AI space is that practitioners are generally right-brained (love math, very technical, logic is king). The audience however can be much more left-brained (creative, emotionally intelligent, visual masterminds). The makes it difficult to "speak the same language". Meet your audience on their level. Be a bit creative, and have fun with the visuals. Check in with them to see if your points are hitting the mark. Good data presented well is magic.

Sometimes, I find toy examples very useful to share in simple terms how the model works. Also, showing that the model successfully predicts some known stilized facts can help audience assess their confidence in the model’s predictive performance. Of course, this is not enough, but a useful start.

I avoid displaying decision trees when discussing Machine Learning, as it might suggest a mere "if-else" statement. Instead, I rely on visuals to juxtapose the new model against the baseline, demonstrating the impact of feature engineering on model performance.

I wholeheartedly concur! Based on my observations, individuals tend to exhibit minimal curiosity regarding the construction and functioning of the model. Rather, their interest lies in understanding the potential impact and the projected ROI resulting from the integration of the model into their system.

For an optimization classical machine learning problem the challenge would be to put a robust explainability framework. One of the framework where trust is built on the optimization model it defining and simulating other "what if" scenarios. The idea is to change the different levers of the objective function and compare the output with the base model. This would helps you to test ideas and make better decisions. "what if" scenarios is like having a crystal ball that shows you the outcomes of different decisions, make informed choices, and navigate your way to the best possible outcome