Discover AWS Certified experts on-demand

A client was looking to migrate a number of their machine learning models to Amazon SageMaker using Jupyter notebooks. They wanted to take advantage of Amazon SageMaker’s automated updates to support batch and real time predictions. To address this request, I created a two-stage proposal.

First, I created an Amazon SageMaker Jupyter notebook for one model. This notebook read the most recent 12 months of data from AWS S3, then cleaned up the data in preparation for training, testing and validation, and trained a binary prediction model using the XGBoost classification algorithm. The notebook then demonstrated how to deploy the model into an endpoint for external apps (such as AWS Lambda), and used Amazon SageMaker batch transform to make bulk predictions.

For the second stage, I video conferenced with the data science team to review the Jupyter notebook. They requested I add additional code to improve the model and address their model input balance concerns. On model performance evaluation, I created a confusion matrix with explanations on interpreting the metrics, and which hyperparameters to further fine-tune in order to reach their business requirements. I also discussed potential new features with the team including feature engineering and extensions of the model for multi class prediction.

Upon the completion of the proposal, the team had enough information, code and knowledge to migrate the remaining models from the existing machine learning solution to Amazon SageMaker. The data science team was very satisfied with the quality and speed of my service as the proposal was completed within one week. I look forward to future collaborations with the team on AWS IQ.