Algorithms used by Autopilot

Hello!

Can somebody tells me what kind of algorithms are supported by Autopilot at the moment? I’ve just found ( deep inside Amazon SageMaker FAQs ) that “Amazon SageMaker Autopilot supports 2 built-in algorithms at launch: XGBoost and Linear Learner.” - Does this mean that only these 2 algorithms are used under the hood to train all kind of models that we can build by means of Autopilot?
And the second related question: Where can I find info - what type of algorithms has been used for my particular task during model building and training? I can find this information only inside notebooks or I can use UI / Reports to get that information?

Hi @Danila,

Welcome to PDS, You can refer to this paper to have an understanding about Amazon SageMaker Autopilot.

Regarding first question, Generally ML algorithms follows no free lunch theorem. If you see kaggle tabular data competitions, most of the participants use XGBoost or similar algorithms like catboost, LightGBM. tree based XGBoost models suffer from higher estimation variance compared to linear models. I guess that’s why the library is also providing linear learner. Even though there are only two algorithms. Each algorithm behaviors can be changed using parameters. Best eg XGBoost classification objective can be based on F1, precision, recall, or accuracy or Tree based or a linear function.

In the paper, it is mentioned that Autopilot produces up to 250 consumable and ready-to-deploy models representing the entire ML pipelines.

Regarding 2nd question, I hope this link might be helpful for you.

Best Regards,
A. Sriharsha

1 Like