Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

SageMaker Canvas end to end machine learning workflow

Setting up

An Amazon SageMaker domain is a centralized place to manage all your Amazon SageMaker environments and resources. A domain acts as a virtual boundary for your work in SageMaker, providing isolation and access control for your machine learning (ML) resources.

To get started with Amazon SageMaker Canvas, you or your administrator must navigate to the SageMaker console and create a Amazon SageMaker domain. A domain has the storage and compute resources needed for you to run SageMaker Canvas. Within the domain, you configure SageMaker Canvas to access your Amazon S3 buckets and deploy models. Use the following procedure to set up a quick domain and create a SageMaker Canvas application.

The preceding procedure used a quick domain set up. You can perform an advanced set up to control all aspects of the account configuration, including permissions, integrations, and encryption. For more information about a custom set up, see Custom setup to Amazon SageMaker.

By default, doing the quick domain set up provides you with permissions to deploy models. If you have custom permissions set up through a standard domain and you need manually grant model deployment permissions, see Permissions management.

Flow creation

Amazon SageMaker Canvas is a machine learning platform that enables users to build, train, and deploy machine learning models without extensive coding or machine learning expertise. One of the powerful features of Amazon SageMaker Canvas is the ability to import and work with large datasets from various sources, such as Amazon S3.

For this tutorial, we're using the NYC taxi dataset to predict the fare amount for each trip using a Amazon SageMaker Canvas Data Wrangler data flow. The following procedure outlines the steps for importing a modified version of the NYC taxi dataset into a data flow.

Data Quality and Insights Report 1 (sample)

After importing a dataset into Amazon SageMaker Canvas, you can generate a Data Quality and Insights report on a sample of the data. Use it to provide valuable insights into the dataset. The report does the following:

It can identify other potential issues that may impact model performance. It also evaluates the predictive power of each feature concerning the target variable, allowing you to identify the most relevant features for problem you're trying to solve.

We can use the insights from the report to predict the fare amount. By specifying the Fare amount column as the target variable and selecting Regression as the problem type, the report will analyze the dataset's suitability for predicting continuous values like fare prices. The report should reveal that features like year and hour_of_day have low predictive power for the chosen target variable, providing you with valuable insights.

Use the following procedure to get a Data Quality and Insights report on a 50,000 row sample from the dataset.

You can review the Data Quality and Insights report on a sample of your data. The report indicates that the year and hour_of_day features are not predictive of the target variable, Fare amount.

At the top of the navigation, choose the name of the data flow to navigate back to it.

Drop year and hour of day

We're using insights from the report to drop the year and hour_of_day columns to streamline the feature space and potentially improve model performance.

Amazon SageMaker Canvas provides a user-friendly interface and tools to perform such data transformations.

Use the following procedure to drop the year and hour_of_day columns from the NYC taxi dataset using the Data Wrangler tool in Amazon SageMaker Canvas.

You can use the preceding procedure as the basis to add all of the other transforms in SageMaker Canvas.

Data Quality and Insights Report 2 (full dataset)

For the previous insights report, we used a sample of the NYC taxi dataset. For our second report, we're running a comprehensive analysis on the entire dataset to identify potential issues impacting model performance.

Use the following procedure to create a Data Quality and Insights report on an entire dataset.

The following is an image from the insights report:

It shows the following issues:

Duplicate rows can lead to data leakage, where the model is exposed to the same data during training and testing. They can lead to overly optimistic performance metrics. Removing duplicate rows ensures that the model is trained on unique instances, reducing the risk of data leakage and improving the model's ability to generalize.

A skewed target variable distribution, in this case, the Fare amount column, can cause imbalanced classes, where the model may become biased towards the majority class. This can lead to poor performance on minority classes, which is particularly problematic in scenarios where accurately predicting rare or underrepresented instances is important.

Addressing data quality issues

To address these issues and prepare the dataset for modeling, you can search for the following transformations and apply them:

If you're not sure about how to apply a transform, see Drop year and hour of day

By addressing these data quality issues and applying appropriate transformations, you can improve the dataset's suitability for modeling.

Verifying data quality and quick model accuracy

After applying the transforms to address data quality issues, such as removing duplicate rows, we create our final Data Quality and Insights report. This report helps verify that the applied transformations resolved the issues and that the dataset is now in a suitable state for modeling.

When reviewing the final Data Quality and Insights report, you should expect to see no major data quality issues flagged. The report should indicate that:

Additionally, the report should provide a quick model score based on a baseline model trained on the transformed dataset. This score serves as an initial indicator of the model's potential accuracy and performance.

Use the following procedure to create the Data Quality and Insights report.

Split the data into training and test sets

To train a model and evaluate its performance, we use the Split data transform to split the data into training and test sets.

By default, SageMaker Canvas uses a Randomized split, but you can also use the following types of splits:

You can change the Split percentage or add splits.

For this tutorial, use all of the default settings in the split. You need to double click on the dataset to view its name. The training dataset has the name Dataset (Train).

Next to the Ordinal encode node apply the Split data transform.

Train model

After you split your data, you can train a model. This model learns from patterns in your data. You can use it to make predictions or uncover insights.

SageMaker Canvas has both quick builds and standard builds. Use a standard build to train best performing model on your data.

Before you start training a model, you must first export the training dataset as a SageMaker Canvas dataset.

After you've created a dataset, you can train a model on the SageMaker Canvas dataset that you've created. For information about training a model, see Build a custom numeric or categorical prediction model.

Evaluate model and make predictions

After training your machine learning model, it's crucial to evaluate its performance to ensure it meets your requirements and performs well on unseen data. Amazon SageMaker Canvas provides a user-friendly interface to assess your model's accuracy, review its predictions, and gain insights into its strengths and weaknesses. You can use the insights to make informed decisions about its deployment and potential areas for improvement.

Use the following procedure to evaluate a model before you deploy it.

You can now view the model evaluation metrics.

After you evaluate the model, you can make predictions on new data. We're using the test dataset that we've created.

To use the test dataset for predictions we need to convert it into a SageMaker Canvas dataset. The SageMaker Canvas dataset is in a format that the model can interpret.

Use the following procedure to create a SageMaker Canvas dataset from the test dataset.

Use the following procedure to make predictions. It assumes that you're still on the Analyze page.

Deploy a model

After you've evaluated your model, you can deploy it to an endpoint. You can submit requests to the endpoint to get predictions.

Use the following procedure to deploy a model. It assumes that you're still on the Predict page.

Cleaning up

You've successfully completed the tutorial. To avoid incurring additional charges, delete the resources that you're not using.

Use the following procedure to delete the endpoint that you created. It assumes that you're still on the Deploy page.

After deleting the deployment, delete the datasets that you've created within SageMaker Canvas. Use the following procedure to delete the datasets.

To avoid incurring additional charges, you must log out of SageMaker Canvas. For more information, see Logging out of Amazon SageMaker Canvas.