New Framework for Markov Chain Choice Models

Traditional choice modeling often assumes that customer transactions are independent, typically drawing from a random utility model. However, real-world customer behavior, especially when observed over time (panel data), often exhibits dependencies among transactions from the same individual. This research addresses this gap by proposing a comprehensive framework for Markov chain (MC) choice models that explicitly accounts for these dependencies. The core contribution is the development of novel expectation-maximization (EM) algorithms for MC parameter estimation. These algorithms uniquely incorporate partial-ordering preference information derived from a customer's historical transaction data, offering a more nuanced understanding of individual choices. Evaluations on synthetic datasets and a real-world sushi dataset demonstrate that these new EM algorithms significantly outperform both traditional EM methods and multinomial-logit-based benchmarks. Beyond estimation, the framework also delves into the computational complexities of conditional choice prediction and personalized assortment optimization, providing valuable insights into these challenging problems and clarifying the computational landscape for choice modeling with panel data.