Trust is a big deal when it comes to machine learning. “Black box” algorithms, concerns about bias and a sense that data scientists may know everything about the data but nothing about the business all undermine trust in machine learning models. Indeed, building machine learning models that can be, will be, trusted is regarded as a critical issue for many data science teams.
Stephen Covey once wrote a famous list about trust – The 13 behaviors of a high trust leader. Five of these behaviors relate very specifically to leadership (talk straight, demonstrate concern, right wrongs, show loyalty, keep commitments), but the others provide a great framework for building trust in your machine learning projects.
- Listen First
Perhaps the most important way data science teams can build trust in their models is to begin not by building a model, not by looking at the data, but by listening to their business partners. Asking businesspeople how they decide, how they would like to decide, and really listening is a great first step. If the business partners feel heard, then they are much more likely to trust the solution the data science team creates. We find it most effective to create a decision model with our business partners to create a shared understanding of the decision and to show that we heard.
- Extend Trust
Data scientists who want their models to be trusted need to extend trust to their business partners. Businesspeople can be wrong – about what moves the dial, about what the threshold should be, about what customer segments the company has – but data scientist should resist the temptation to assume that they are wrong. It’s easy to say that the data can answer all these questions and to give the impression that the businesspeople’s expertise is not trusted. Doing so will make it harder to get to trust any machine learning model the team delivers. Begin by extending trust – by assuming that the decision model they tell you to build is correct.
- Clarify Expectations
Before you build a machine learning model to influence the decision your business partners have described, make sure you clarify your expectations. If your model fits easily into the current decision, improving its accuracy, say so. If your model is more likely to disrupt the current approach to decision-making, require significant organizational change then say that. Use the decision model to clarify the expectations you have for the use of the models you are building. And as you build them, as you find out what you can and can’t predict, keep using the decision model to align everyone’s expectations.
- Confront Reality
Do not pretend that the organization will change the way it makes decisions just because your team tells it to. The reality of many decisions is that they are constrained by regulations, driven by policies and motivated by goals and objectives. Your model may be able to improve the decision-making “in theory” but to improve it “in practice” you must confront reality. A shared understanding of the decision-making approach, in the form of a decision model, can help.
- Create Transparency
The importance of transparency in machine learning – explainable AI or reason codes for a model – is well-established. To drive real success, however, you need also to create transparency about how your model will be used. It is critical to know the score your model creates and why that is the score. You must also be able to show how that score was used to change decision-making and impacted business results. A clear, shared understanding of the decision-making that wraps around the machine learning model is key. Use the decision model to maximize transparency and add transparency to your analytic too.
- Deliver Results
Deliver business results. Lift and model accuracy are not results. Improved business outcomes are results. Your job is to improve business results, which means improving the way decisions are made, not just to produce an analytic or machine learning model. Make sure you can deliver the result your business partners care about and don’t talk only about how good your model is. Refer back to the decision model to put your analytic results in a business context.
- Practice Accountability
Be accountable to your business partners. Remember that they have other things to do besides work with you, other projects they must support. Remember that they have business objectives to meet and that your model needs to help them to meet those objectives. Be accountable to the business problem, not just to the analytic solution.
- Get Better
One of the most important lessons we have learned automating decisions and deploying machine learning and other analytics is that continuous improvement is key. Don’t try and develop a perfect model, get a minimum viable analytic product built and in use to see how it affects business outcomes. Capture data about how decisions were made, how they used (or didn’t use your analytic) and how well that worked out. Don’t regard an analytic project as a one-time effort. Don’t ride off into the sunset when you have built a model. Focus on how you can help the business get better now and in the future.
If you want to learn more about decision modeling and use it to build trust in your machine learning projects, contact us.