The mixed media modeling (MMM) analysis in marketing is a machine learning technique that combines information from various sources such as TV ads, online ads, and social media to measure the impact of marketing and advertising campaigns. By using these techniques, companies can make smarter decisions about where to invest their advertising money, helping them achieve the best return on investment. This approach is like having a treasure map that guides you towards the most valuable marketing strategies.
A significant challenge for marketing analysis teams is the high computing requirements needed to run these models. Popular libraries like LigeroMMM, Robyny Marketing, and PyMC are not designed to scale horizontally, causing delays in understanding and optimizing marketing strategies, especially when analyzing granular demographic and geographic regions.
This blog post demonstrates how to accelerate MMM modeling jobs using AWS Batch. By leveraging LigeroMMM as an open-source library, AWS Batch helps address two key challenges: providing teams access to larger GPU and CPU computing resources, and efficiently provisioning and managing infrastructure to reduce the risk of overspending. This results in a significant reduction in model training time, enabling faster decision-making and optimization of marketing strategies.
The architecture of the sample application integrates AWS Batch with a web front end, allowing users to easily submit training jobs for MMM. The frontend web application calls an API in the processing layer, which then uses AWS Batch to run training jobs. This setup ensures that resources are automatically provisioned and managed, reducing costs and increasing efficiency.
Furthermore, the post provides in-depth analysis of the architecture components, including the web interface, data lake for storage, processing layer, and the LightweightMMM framework. Parameters influencing model training runtime, such as historical data, granularity, number of geographies, marketing channels, and chains, are discussed in detail.
Tips on maximizing GPU memory, setting up CUDA and cuDNN in AWS Batch, and cost-time analysis for running different model configurations are also covered. The analysis shows that using GPU instances led to significant improvements in training time, with instances such as P5 GPUs with NVIDIA H100 reducing training time from 1.5 days to 45 minutes for a moderate cost increase.
In conclusion, the post highlights the benefits of using AWS Batch for faster MMM modeling with Amazon EC2 instances, enabling data scientists to iterate more quickly and create better models more efficiently. To get started with running mixed media models on AWS Batch, users can refer to the GitHub repository provided.
Article Source
https://aws.amazon.com/blogs/hpc/how-vertical-scaling-and-gpus-can-accelerate-mixed-media-modelling-for-marketing-analytics/