Robust estimation of treatment effects in high-dimensional heterogenous data with application to e-commerce

点击翻译按钮获取中文摘要

Granify is an Edmonton-based company that uses artificial intelligence, behavioral messaging and predictive**analytics to optimize e-commerce conversion rates and revenues. During a customer's visit, Granify monitors**various attributes per second such as scroll speed, products and images viewed, mouse movements, and**hesitations, to fully understand the customer's digital behavior. If Granify determines a specific customer's**objection is likely to prevent them from purchasing, it will automatically introduce a message or stimuli to**alleviate their concern and increase their chance of purchasing.To quantify how customers respond to the**message or stimuli introduced by Granify, we need to estimate the treatment effect of the Granify intervention.**A standard A/B test is usually conducted by randomly segmenting customer sessions into groups, both with and**without Granify intervention (treatment and control group, respectively).Currently, average treatment effect**(ATE), the difference of the average money spent per session between treatment group and control group, is**used to measure the effect of implementing Granify. However, ATE only provides a partial view over the**effects of the treatment since each customer might respond to the treatment in different ways. Underlying this**average effect for a sample may be substantial variation in how particular customers respond to treatments.**This variation may reveal how the effect of the Granify intervention depends on customers' characteristics, and**it is useful to provide guidance on how to optimally administer treatments. Furthermore, ATE is often sensitive**to outliers and skewness of the outcome distribution. The goal of this research is twofold: first, to test and**estimate heterogeneous treatment effects of Granify intervention; and second, to provide a robust location**estimator and corresponding confidence interval for the treatment effect.