Acxiom Uses Amazon SageMaker for Propensity Scoring 3 Trillion Records

Acxiom is a customer intelligence company that offers data-driven insights to marketers for personalized advertising campaigns. A global data leader with thousands of data attributes in more than 30 countries, Acxiom manages a customer intelligence platform at scale.

Acxiom uses predictive modeling to generate audience propensity scores used to help marketers understand which products consumers are shopping for, whether consumers are interested in purchasing a product, and which content consumers are watching or streaming. To accomplish this, Acxiom has a catalog of 5,000 audience propensity models generating a total of 3 trillion propensity scores across a range of categories including retail, automotive, CPG, healthcare, finance, and insurance. Prior to running on Amazon Web Services (AWS), Acxiom used an on-premises Apache Hadoop cluster that ran monthly and took 11–15 days to score all 5,000 models, which consumed upwards of 90 percent of their processing capacity during that period and created additional operational challenges around scalability and infrastructure management.

Acxiom used Amazon SageMaker to build an end-to-end machine learning (ML) pipeline for audience propensity scoring. Acxiom’s inference architecture pipeline uses Amazon DynamoDB to export, analyze, and stream data on Amazon Simple Storage Service (Amazon S3), and SageMaker processing to build, train, and deploy customer propensity models. As a result of moving the workload to AWS, Acxiom reduced inference time by 73 percent and reduced total cost of ownership by 61 percent compared to its previous architecture.

“[Amazon SageMaker is] a one-stop shop solution for model training, model deployment, model inferencing, and quality assurance,” says Varadarajan Srinivasan, senior director of data science and machine learning engineering at Acxiom. “And it offers significant financial flexibility because it's pay for use.” Using AWS, Acxiom can quickly build and deliver audiences at scale, and “if there is a new consumer trend emerging, we can capture the data, model it out, create a new product offering, and ship it to our clients.”

Watch “How Acxiom used Amazon SageMaker for propensity scoring for 5,000+ models and 3 trillion records” from AWS re:Invent 2021.