Machine Learning for Business Analytics, 2nd Edition

14 INTERVENTIONS: EXPERIMENTS, UPLIFT MODELS, AND REINFORCEMENT LEARNING

14.1 INTRODUCTION

In this chapter, we describe a third paradigm of machine learning, different from supervised and unsupervised learning, that deals with interventions and feedback. Data used in supervised and unsupervised learning are typically observational data, being passively collected about the entities of interest (customers, households, transactions, flights, etc.). In contrast, in this chapter, we discuss experimental data, resulting from applying interventions to the entities of interest, and measuring the outcomes of those interventions. We start with the simplest form of intervention—the A/B test, which is a randomized experiment for testing the causal effect of a treatment or intervention on outcomes of interest. A/B tests are routinely used by internet platforms such as Google, Microsoft, and Uber for testing new features. We then describe uplift modeling, a method that combines A/B testing with supervised learning for providing customized targeting. Uplift modeling is commonly used in direct marketing and in political analytics. Finally, we describe reinforcement learning, a general machine learning methodology used in personalization, where the algorithm (or agent) learns the best treatment assignment policy by interacting with experimental units through dynamic treatment assignment and gathering their feedback.

Interventions in JMP: A/B testing is available in JMP. For uplift modeling, ...