Booking.com: Converted data is enough for most promotion uplift modeling problems in E-commerce
Promotions are crucial in e-commerce for increasing user engagement and building customer loyalty. These incentives naturally raise causal questions, as they function as interventions that alter the business’s status quo. Effectively managing the cost-effectiveness of promotions is crucial for creating campaigns that benefit both customers and businesses. This involves addressing a dual causal inference challenge: balancing the uplift in rewards, such as conversion and revenue, against the corresponding increase in costs. In e-commerce, promotional costs and rewards are typically incurred only upon conversion, resulting in zero cost and reward for non-conversions. Given that conversion rates are often below 10%, the data is predominantly zero-inflated, with over 90% of the observations providing limited information about the targets of interest. In this presentation, we demonstrate that in scenarios with triggered costs and rewards, effective promotion modeling can be achieved by applying uplift modeling techniques exclusively to converted data from randomized experiments (e.g., A/B tests). This approach enables the accurate calculation of key metrics, such as lift, relative metrics (revenue, utility, etc.), and Return on Investment (ROI), while also reducing training times by more than 90% without compromising model performance.
-
Hugo Proença Senior Machine Learning Scientist, Booking.com