There are increasing demands for “causal ML models” of the agent behaviors, which enable us to unbox the complex black-box models and make inferences or do proper counterfactual simulations. Many applications (especially in Marketing) intrinsically call for measurement of the causal impact of the product, pricing, and promotion decisions beyond predictions, either from observational data or combinations of experimental and observational data.
Discrete choice modeling of agent behaviors is a generative modeling framework, created by an Economist, Daniel McFadden (Nobel Prize, 2000). This has been a work-horse model in Economics, Marketing Science, and Operation Research. However, this modeling framework is less known to ML/AI researchers outside of Computational Social Science. In this talk, I will introduce discrete choice models of agent behaviors with a focus on consumer demand modeling. I will talk about two different ways of modeling consumer heterogeneity: discrete vs. continuous. In addition, how this individual-level model (i.e. varying parameters at the individual level) can be estimated by using simulated individuals when you only have aggregate sales data is also discussed. A dynamic version of this model is related to reinforcement learning, and I will discuss this linkage. Finally, an extension of this model to consumer online search behaviors and a neural network representation of discrete choice models will be discussed.
We are looking for passionate people willing to cultivate and inspire the next generation of leaders in tech, business, and data science. If you are one of them get in touch with us!