A plea for simple, easy-to-understand model cards

Engineers?at Google ?proposed the concept of the model card?in 2018 ?to increase ML transparency and to create a way to communicate usage, ethics-informed evaluation, and limitations. Many companies have tried their hand at creating their own version of the model card, but it is still to be seen whether any single model card will rise as the unified standard for model documentation.

In my opinion, the reason there isn’t a unified standard for the model card is simple… model cards aren’t simple!

It makes sense… models can be complicated; but this is part of the problem. For so long, we have relied on data scientists and developers at our companies to understand the inner workings of the models an organization develops. But here’s the thing… every day, more and more AI systems are being used to make critical decisions about us. This means that everyone should have the right to understand what these models are and how they work.

From an organizational perspective, model understanding is key because ultimately, we are responsible for the AI systems that our organizations create and deploy. This means that our managers and executives also need to understand these AI systems, even if it’s just a high-level understanding. While it’s a hard ask, we need to create model documentation that everyone can understand.

When SAS started model card development, we made it our primary objective for the model card to be simple, yet robust. And that is exactly what we did .

The landing-page for the SAS model card (Overview) helps an executive answer the question, “Is this model any good?” We provide a simple “Pass/Fail” message about metrics such as accuracy, generalizability (i.e., can this model work well on data it’s never seen before), fairness, and model drift. We don’t expect the user to know what any of those metrics mean, but if they see a big, scary, red “Fail” on the screen, that will probably answer their question about whether the model is any good. At this point, they’d hand the baton to their technical teams to resolve that “Fail” notification.

Even though the SAS model card was designed with executives in mind, we also designed it with others in mind:

• The “Model Usage” section is intended for those that will ultimately use the model. It explains how a model should be used, and more importantly, how it shouldn’t be used. This section allows model developers to document any ethical considerations and limitations users need to be aware of.
• The “Data Summary” section is intended for business analysts and data engineers. It provides a high-level look into the dataset that was used to train the model.
• The “Model Summary” section is the most technical and is intended for data scientists. It shows the model type, the outcome/target variable, and a slew of other metrics a user can choose from to evaluate the performance of the model. This section also includes a fairness assessment, which will surface any differences in model performance by variables deemed most important for fairness evaluation.
• The “Model Audit” section is a double-click of the “Pass/Fail” information shown in the Overview section and is intended for data scientists and model engineers. It shows the metrics that have been chosen to be evaluated over time, their alert thresholds, and whether the metrics have met/surpassed the acceptable thresholds.

One more important thing to mention… without an automated process to create model cards, this takes a lot of time and effort. Organizations are continually asking their developers and data scientists to add more steps to their processes; it’s going to be a hard sell to convince your data scientists to manually gather all of the information required to populate a model cards.

This was another critical requirement in the development of the SAS model card… make it easy to create! And this is exactly what we did. SAS model cards are automatically generated upon model registration! So, data scientists, you can breathe a sigh of relief… phew!

Whether we will ever have a unified standard for model cards remains to be seen, but we need to do everything in our power to promote responsible innovation. So please… if you take nothing else from this article, hear this: please make model cards easy to create, and more importantly, easy to understand!

-

Allie DeLonay is a Senior Data Scientist with SAS’ Data Ethics Practice (DEP). She works with research & development teams to brainstorm and implement the trustworthy and ethical AI features required for SAS to substantiate their claim as a leader in responsible innovation.