Home > Articles > Building Trustworthy Models

Building Trustworthy Models

Machine learning has a trust problem. Discussions about the role that algorithms play in our lives have become national (if not global), with some raising important and legitimate questions about the biases inherent in these algorithms. In this environment, we wonder: what would it take for a model to be worthy of our trust?

I recently read an illuminating piece by David Spiegelhalter called "Should We Trust Algorithms?". In it, he identifies the difference between trustworthy claims about a system and trustworthy claims made by a system. His article spends more time on the former than the latter, so I've written this article to elaborate on ways our models can make more trustworthy claims.


Building trust with users is essential for a few reasons. First, we want our products to be used. If a user doesn't trust the predictions made by my model, she is less likely to follow its advice. Worse, without adequately communicating uncertainty, we may actively anger her. Suppose a model predicts this user's house will sell for $300,000, but it ends up selling for $290,000. It's difficult to fault her for being upset at this $10,000 difference.

By contrast, if we predicted a range of possible sale values, the user would have better expectations going in and a better experience with our product.

Ethics provide a second reason that building trust is paramount. It is unethical to present estimates without a sense for their uncertainty. A common machine learning approach is to build some classification or regression model for a problem. These models typically output a single predicted value: "this flower is a setosa", "this house is worth $300K", or "this image has an airplane in it". These statements imply a level of certainty that may be unwarranted by the data, and we must be very careful to place them into context to avoid dishonesty.

Finally, trustworthy models are just good business. If we hide uncertainty with overly-precise point estimates, we are likely to make bad decisions. Annie Duke writes:

A great decision is the result of a good process, and that process must include an attempt to accurately represent our own state of knowledge. That state of knowledge, in turn, is some variation of ā€œIā€™m not sure.ā€

Trustworthy models in practice

But building trust is hard. We cannot simply tell our users to "trust the algorithm" and expect them to do so. Instead, Spiegelhalter argues, we should put our efforts into building models that are worthy of trust. When we relate to other humans, we understand that we must demonstrate that we are worthy of trust before we will be trusted. The same holds true for models.

A coworker of mine once asserted that trustworthy models provide at least three things:

  1. Prediction -- An estimate for some unknown value
  2. Confidence -- A description of how uncertain the model is about the prediction
  3. Explanation -- The reasoning for which a model made its prediction

As an example, consider a doctor. If your doctor told you that your arm needed to be amputated, you'd never let them do it just based on that recommendation! You would ask for some justification first. She would explain that an infection in your arm could be lethal if it spreads. In this way, she builds trust with you.

These techniques which come as second nature to humans are not as automatic for machines. We often stop short of developing models that are truly worthy of our users' trust.

Here's an example of what the output of a trustworthy model could look like:

  1. Prediction: My best guess at the sale price of this home is $324,000
  2. Confidence: A 95% confidence interval on that number is ($315K, $333K)
  3. Explanation: The home is large, but the fact that it's on a corner brings its value down.

This is a massive improvement over a single number! Not only is this a more honest statement to make, it helps users understand why the model gave its prediction. In turn, this helps users gain trust in the system and leads to better outcomes.

We have a responsibility as model builders to represent our work with integrity. Shipping a model that is implicitly overconfident is bad for our users and our businesses. Instead, we should develop models that are truly worthy of trust.