Ax - Multi-Objective Optimization for Marketing Campaigns

Summary¶

This project demonstrates the use of the Ax (Adaptive Experimentation) library to leverage the Bayesian Optimization technique to optimize the parameters of a marketing campaign.

The content of this project includes the description of the Ax library, the Bayesian Optimization algorithms, and the implementation of this method to optimize the hyperparameters of a marketing campaign in a real-time bidding (RTB) scenario.

Getting Started¶

Prerequisites¶

The project is designed to be run in a Docker container. As long as Docker is available, the project can be run locally without the need to install any dependencies.

Installation and Docker Setup¶

1. Build the Docker image

   > ./docker_build.sh

2. Run the Docker container

   > ./docker_bash.sh

3. Open the Jupyter notebook

   > ./data/run_jupyter.sh

4. Open in browser

   • Go to http://localhost:8888 in your web browser

Structure of the Project¶

Auxiliary Files¶

• docker_build.sh
  • Builds the Docker image for the project
• docker_bash.sh
  • Runs the Docker container and opens the Jupyter notebook
• docker_clean.sh
  • Removes the Docker image from the local system
• docker_exec.sh
  • Executes a command in the Docker container
• docker_push.sh
  • Pushes the Docker image to the Docker Hub
• install_jupyter_extensions.sh
  • Installs the Jupyter extensions
• run_jupyter.sh
  • Runs the Jupyter notebook
• version.sh
  • Displays the versions of the installed software and packages

Description of the API¶

The API documentation as well as the code to demonstrate the use of this API is available in 1. Ax - Multi-Objective Optimization for Marketing Campaigns.

Real-Time Bidding Algorithms with Bayesian Optimization¶

1. Ax - Multi-Objective Optimization for Marketing Campaigns demonstrates the use of the Ax library to find the optimal hyperparameters for a bidding strategy in a real-time bidding campaign.

DSP Simulation¶

Bayesian Optimization is used to find the optimal hyperparameters for the bidding strategy. This strategy requires the prediction of the Click-Through Rate (CTR) usually done by a marketing DSP platform.

The prediction of the CTR is out of the scope of this project. In order to demonstrate how the hyperparameter optimization works, a simulation of a CTR prediction model is used.

The creation of this predictive model is described in dsp_pctr_prediction_model.ipynb.

Note: To save time during the execution of this tutorial, the predicted CTR has been appended to the dataset. The bidding simulation will know beforehand what’s the CTR a machine learning model would have predicted.

Multi-Armed Bandits¶

A second example of Bayesian Optimization is the use of the Ax library to find the optimal bandit strategy to maximize the reward in a multi-armed bandit problem. It’s also described in 1. Ax - Multi-Objective Optimization for Marketing Campaigns.