ExProCO: an explainable probabilistic campaign optimizer for ecommerce advertising

Abstract

Optimizing advertising (Ad) campaigns on e-commerce platforms is a complex task that extends beyond identifying correlations between budgets and bids or predicting metrics such as impression share and Return on Advertising Spend (ROAS). Effective Ad optimization involves addressing the auction-based nature of e-commerce advertising, which inherently requires a probabilistic approach to account for uncertainties and dynamic conditions. Consequently, a novel approach that combines probability analysis with decision tree modeling is proposed. The proposed model ExProCO is built on a Greedy-Modal Tree (GMT), offering a well-interpreted strategy for extracting useful information from complex, high-dimensional data at the campaign, keyword, and product levels. Using GMT, ExProCO minimizes Decision Tree instability and provides explainability. The Joint Probability Distribution (JPD) Model uses daily campaign data to explain how variables like bids and budget affect campaign outcomes. The ExProCO model outperforms other Machine Learning models in a benchmark comparison using financial data from a leading e-commerce firm. Achieving 75.4% accuracy with just 11.3 nodes, ExProCO excels in interpretability, noise resilience, and overfitting reduction, significantly improving campaign profitability and ad spend optimization.