A1090
Title: Conjugating variational inference for mixed logit models
Authors: Weiben Zhang - University of Melbourne (Australia) [presenting]
Ruben Loaiza-Maya - Monash University (Australia)
Michael Stanley Smith - Melbourne Business School (Australia)
Ole Maneesoonthorn - Monash University (Australia)
Abstract: Contemporary choice models often account for heterogeneous taste variation, resulting in models with many latent variables. Traditional Markov chain Monte Carlo (MCMC) methods become impractical for models with large datasets or high-dimensional latent variables. For cases where latent variables are easy to sample, such as the probit model, hybrid variational inference (HVI), which combines Monte Carlo sampling for latent variables with a variational approximation for model parameters, offers a fast and accurate alternative. A conjugating variational inference (CVI) method is introduced, designed for cases where sampling latent variables is challenging, such as in mixed logit models. CVI is applied to three types of choice models with latent variables: the mixed multinomial logit (MMNL) model, the mixed nested logit (MixNL) model, and the mixed bundle choice model. Using both simulated data and retail datasets, the computational speed and predictive accuracy of the proposed method are compared against several benchmark approaches.
