In recent years, AWS customers have been running machine learning on an increasing variety of datasets and data sources. Since a large percentage of organizational data is stored in relational databases such as Amazon Aurora, there's a common need to make this relational data available for training ML models, and to use ML models to make predictions in database-based applications.
Here, we show how to easily extract your production data from Amazon Aurora, train an ML model, then integrate the model inferences back into your production database and applications. We extend a popular ML use case, predicting customer churn, and show how to achieve the real business goal – preventing the customer churn.
We explore the explainability of the XGBoost model, and use that information to provide Marketing with a starting point for developing targeted incentives. We then show how to use a single SQL statement in Amazon Aurora to combine data from its tables, an Amazon SageMaker inference on the customer churn likelihood, and an Amazon Comprehend assessment of a customers message sentiment. This response can be fed in real time through a Python function that chooses the appropriate incentive to recommend, based on rules from Marketing, and returned along with the customer's basic information. Perfect for use by a customer call center!
The blog post associated with this repo is located here. The blog post includes a button that executes the CloudFormation template in region us-east-1. The blog builds the infrastructure shown in the diagram, and walks through the associated steps.
To reuse or modify the code, clone this repository. Place the resources in your bucket
on Amazon S3. Update the CloudFormation template (see SageMakerNotebookInstanceLifecycleConfig) to point to the correct bucket
name to use when loading the resources, as the template currently has the
locations of the notebooks hardcoded in it. You will need to modify these
locations to point to your own S3 bucket.
To get the notebooks as HTML files with code and output use the following links: part1, part2, part3.
This sample code is made available under the MIT-0 license. See the LICENSE file.