Budget Allocation at Scale: Building an ML Optimization System for Paid Marketing

This post details the architecture and implementation of Atlassian's two-stage ML system for optimizing paid marketing budget allocation. It introduces the payout model, which uses LightGBM and captures spend diminishing returns, and the budget optimizer, which evaluates Bayesian Optimization and the Cross-Entropy method. The post also describes the online experiments conducted using a pre-and-post design and highlights the achieved double-digit percent reduction in cost per business metric. Future work includes improving extrapolation handling, incorporating exploration and exploitation strategies, and enriching features.