Explainable AI for Transparent Decisions with DataRobot AI Platform

For many years, there was a lot of mystery around AI. When we can’t understand something, we struggle both to explain it and trust it. But as we see a rise in AI technologies, we need to challenge systems to be sure if it is trustworthy. Is it reliable or not? Are decisions fair for consumers or do they benefit businesses more? 

At the same time, a McKinsey report notes that many organizations get tremendous ROI from AI investments in marketing, service optimization, demand forecasting, and other parts of their businesses (McKinsey, The State of AI in 2021). So, how can we unlock the value of AI without making huge sacrifices to our business?

Explainability in DataRobot AI Platform  

In DataRobot, we are trying to bridge the gap between model development and business decisions while maximizing transparency at every step of the ML lifecycle—from the moment you put your dataset to the moment you make an important decision.

Before jumping into the technical details, let’s also look at the principles of technical capabilities:

• Transparency and Explainability
• Fairness 
• Governance and Risk Management 
• Privacy and Security

Each of these components is critical. In particular, I would like to focus on explainability in this blog. I believe transparency and explainability are a foundation for trust. Our team worked tirelessly to make it easy to understand how an AI system works at every step of the journey. 

So, let’s look under the hood of the DataRobot AI platform.

Understand Data and Model 

The great thing about DataRobot Explainable AI is that it spans across the entire platform. You can understand the model’s behavior and how features affect it with different explantation techniques. For example, I took a public dataset from fueleconomy.gov that features results from vehicle testing done at the EPA National Vehicle and Fuel Emissions Laboratory and by vehicle manufacturers.  

I just dropped the dataset in the platform, and after a quick Exploratory Data Analysis, I could see what was in my dataset. Are there any data quality issues flagged? 

No significant issues are spotlighted, so let’s move ahead and build models. 

Now let’s look at feature impact and effects. 

Feature Impact tells you which features have the most significant influence on the model. Feature Effects tell you exactly what effect changing a part will have on the model. Here’s the example below.

And the cool thing about these both visualizations is that you can access them as an API code or export. So, it gives you full flexibility to leverage these built-in visualizations in a comfortable way. 

Decisions that You Can Explain

It took me several minutes to run Autopilot to get a list of models for consideration. But let’s look at what the model does. Prediction Explanations tell you which features and values contributed to an individual prediction and their impact. 

It helps to understand why a model made a particular prediction so that you can then validate whether the prediction makes sense. It’s crucial in cases where a human operator needs to evaluate a model decision, and a model builder must confirm that the model works as expected. 

Deeper Dive into Your Models and Compliance Documentation 

In addition to visualizations that I already shared, DataRobot offers specialized explainability features for unique model types and complex datasets. Activation Maps and Image Embeddings help you understand visual data better. Cluster Insights identifies clusters and shows their feature makeup.

With regulations across various industries, the pressures on teams to deliver compliant-ready AI is greater than ever. DataRobot’s automatic compliance documentation allows you to create custom reports with just a few clicks, allowing your team to spend more time on the projects that excite them and deliver value.  

When we feel comfortable with the model, the next step is to ensure that it gets productionalized and your organization can benefit from predictions. 

Continuous Trust and Explainability 

Since I am not a data scientist or IT specialist, I like that I can deploy a model with just a few clicks, and most importantly, that other folks can leverage the model built. But what happens to this model after one month or several months? There are always things that are out of our control. COVID-19, geopolitical, and economic changes taught us that the model could fail overnight. 

Again, explainability and transparency resolve this issue. We combined continuous retraining with comprehensive built-in monitoring reporting to ensure that you have complete visibility and a top-performing model in production—service health, data drift, accuracy, and deployment reports. Data Drift allows you to see if the model’s predictions have changed since training and if the data used for scoring differs from the data used for training. Accuracy enables you to dive into the model’s accuracy over time. Finally, Service Health provides information on the model’s performance from an IT perspective.

Do you trust your model and the decision you made for your business based on this model?Think about what brings you confidence and what you can do today to make better predictions for your organization. With DataRobot Explainable AI, you have full transparency into your AI solution at all stages of the process for any user.

About the author

Yulia Shcherbachova

Director, Product Marketing at DataRobot

A marketing expert with 10 years of experience in the tech space. One of the early DataRobot employees. Yulia has been working on various company strategic initiatives across different business functions to drive the adoption, product enablement, and marketing campaigns to establish DataRobot presence on the global market.

Meet Yulia Shcherbachova