
I’m diverging from the previous articles in the series. I’m going to review two tools that are heads and shoulders above the others. The design and beautiful visualizations do not come cheap. That doesn’t mean we can’t admire them and use them as a bar to which we strive. I will start with DataRobot. It’s an enterprise tool that you may find yourself having access to through work or school.
Why DataRobot?
I have experience using this tool and love it for the business cases for which I use it. My business case is to have a straightforward interface for a non-data scientist to run and deploy models in an automated way. DataRobot adds new features on a regular cadence, each built nicely within the existing user experience. I could go on about the benefits, but I will control my inner fan-girl.
To keep things even with the other tools, I will focus on the most basic tasks to run a simple .csv file with Automl without any manual interventions or hyper-parameter tuning.
The setup and cost
Straight up, DataRobot is outside of the budget range of the individual data scientist. The implementation and cost are definitely in the realm of businesses. AWS Marketplace offers a one-year subscription for $98,000. Pocket change, I’m sure. But if you use AWS govCloud, it is $9.33/hr (it varies). Interesting.
The Data
To keep parity across the tools in this series, I will stick to the Kaggle training file. Contradictory, My Dear Watson. Detecting contradiction and entailment in the multilingual text using TPUs. In this Getting Started Competition, we’re classifying pairs of sentences (consisting of a premise and a hypothesis) into three categories – entailment, contradiction, or neutral.
6 Columns x 13k+ rows – Stanford NLP documentation
- id
- premise
- hypothesis
- lang_abv
- language
- label
Loading the data
You create a project by uploading a dataset. This interface is where you begin.

After the data is loaded, there are opportunities to change datatypes or remove features. There are some data distribution data. A bonus is that there are warnings if there might be data leakage. If data leakage is detected, DataRobot removes that feature from the final training dataset.


Training your model
Once you choose your target, you hit the big Start button with Modeling Mode set to AutoPilot. When you do that, you will see progress on the right side. As models are trained, they become available on the leaderboard as they complete.
One good thing about having access to the early model results is that you can review for significant issues. Many times some data issues become glaringly apparent with the Insights, and I could halt the process and try again. This quick and easy review helps with rapid iteration.
Evaluate Training Results
The leaderboard begins to fill with the completed models. You can choose several valid metrics in the dropdown. There are also some helpful tags to let you know WHY the leaders are up at the top.

You can compare the models against each other.

One tab I use often is speed versus accuracy. There are times when you are scoring millions of records when speed trumps accuracy if the accuracy drop is minor.


The Insights tab is handy. You can quickly see if one of your features is popping. It’s up to your business expertise to know if that’s appropriate or not. This tab is where I find data issues early in the autoML model training. If I see something that doesn’t seem correct, I can iterate faster than waiting for the entire process to finish.

DataRobot model explainability is the best of the tools I have reviewed so far. Each prediction is assigned which features influenced the final score, indicating not only strength but also direction.

Not to be underestimated, documentation can be a real drain on your time. For this simple dataset, DataRobot generates a 7000+ word document with all of the charts, model parameters, and challenger model details. This documentation is a unique feature that I haven’t found in any other tools, though I have asked for it when asked. All done with a single click.


Conclusions
To loosely compare results between tools, I reran the dataset in classification mode. The metrics are just slightly higher than Azure. For the most part, the model results are similar.
For my business case, this is the top of the pile so far. Head-to-head in image processing or time-series may provide different results. That would be a challenge for another series.
The ease of use, visualizations, access to challenger model details, model explainability, and the automated documentation stand out from the others. Of course, you are paying dearly for this.
Next, I will show you H2O.ai Driverless AI. In my opinion, they are the closest comparison to DataRobot at this time. They have gone to great lengths to get top Data Visualization designers on the project so I’m expecting great things.
If you missed one of the articles in the series, I have posted them below.
Is AWS Sagemaker Studio Autopilot ready for prime-time?