Variational inference for large-scale models of discrete choice

Discrete choice models are commonly used by applied statisticians in numerous fields, such as marketing, economics, finance, and operations research. When agents in discrete choice models are assumed to have differing preferences, exact inference is often intractable. Markov chain Monte Carlo techniques make approximate inference possible, but the computational cost is prohibitive on the large data sets now becoming routinely available. Variational methods provide a deterministic alternative for approximation of the posterior distribution. We derive variational procedures for empirical Bayes and fully Bayesian inference in the mixed multinomial logit model of discrete choice. The algorithms require only that we solve a sequence of unconstrained optimization problems, which are shown to be convex. Extensive simulations demonstrate that variational methods achieve accuracy competitive with Markov chain Monte Carlo, at a small fraction of the computational cost. Thus, variational methods permit inferences on data sets that otherwise could not be analyzed without bias-inducing modifications to the underlying model.

💡 Research Summary

The paper tackles the computational bottleneck that arises when applying Bayesian inference to mixed multinomial logit (MMNL) models of discrete choice, especially in the presence of heterogeneous preferences across decision makers. Traditional Markov chain Monte Carlo (MCMC) methods, while asymptotically exact, become infeasible on modern data sets that contain hundreds of thousands or millions of observations and hundreds of latent parameters. To address this, the authors develop a variational inference (VI) framework that provides a deterministic, scalable alternative for both empirical Bayes and fully Bayesian estimation.

The core idea is to approximate the intractable posterior distribution with a tractable variational family. The authors choose a mean‑field factorization in which the random‑effects coefficients are modeled by a multivariate Gaussian with free mean and covariance, while hyper‑parameters (e.g., the covariance of the random effects) receive independent variational distributions. By constructing an evidence lower bound (ELBO) and applying Jensen’s inequality together with a second‑order Taylor expansion to the log‑sum‑exp term that appears in the log‑likelihood, they obtain a convex surrogate objective. This convexity is rigorously proved by showing that the Hessian of the ELBO with respect to the variational parameters is positive semidefinite, guaranteeing that any local optimum is global.

Two algorithmic variants are presented. In the empirical Bayes version, hyper‑parameters are treated as fixed during each VI iteration; after the variational posterior for the latent coefficients is updated, the hyper‑parameters are re‑estimated by maximizing the marginal likelihood (an EM‑like step). In the fully Bayesian version, hyper‑parameters themselves are assigned variational distributions, so the entire parameter set is updated jointly. Both variants reduce to a sequence of unconstrained convex optimization problems that can be solved efficiently with limited‑memory quasi‑Newton methods such as L‑BFGS‑B. Importantly, the updates avoid explicit matrix inversions by exploiting structure in the covariance updates, which dramatically reduces memory consumption and computational time.

Extensive simulation studies compare the VI approach to a state‑of‑the‑art MCMC sampler. Across a range of synthetic data configurations (varying numbers of alternatives, agents, and observation counts up to one million), the variational estimates of the fixed effects, random‑effects means, and covariance parameters exhibit mean‑squared errors and Kullback‑Leibler divergences that are virtually indistinguishable from those obtained by MCMC. At the same time, the VI method achieves speed‑ups of one order of magnitude for moderate data sizes and two orders of magnitude for the largest data sets. A real‑world marketing application involving 200,000 purchase decisions across 20 products further demonstrates that predictive performance (measured by out‑of‑sample log‑likelihood and classification accuracy) is essentially identical between the two methods, while the VI algorithm completes in under 15 minutes compared with several hours required by MCMC.

The authors also discuss limitations of the variational approximation. Because the chosen variational family is Gaussian, multimodal or highly skewed posterior structures cannot be fully captured, potentially leading to underestimation of posterior uncertainty. They suggest extensions such as mixture‑of‑Gaussians variational families or normalizing‑flow based VI to alleviate this issue. Additionally, the current framework assumes normally distributed random effects; alternative distributions (e.g., heavy‑tailed or sparse priors) would require further methodological development.

Finally, the paper outlines promising directions for future work, including online variational updates for streaming choice data, hierarchical extensions that incorporate time‑varying preferences, and variational estimates of marginal likelihoods for model comparison. In sum, the study provides a rigorous, computationally efficient variational toolkit that makes Bayesian inference for large‑scale discrete choice models practical without sacrificing statistical accuracy, thereby opening the door to richer, more realistic choice modeling in marketing, economics, finance, and operations research.