Diffusion Models Optimize Assortments for Revenue Management

Assortment optimization is a critical challenge in revenue management, traditionally tackled using parametric choice models like the multinomial logit (MNL). While these models offer tractability, their effectiveness is often limited by how accurately they capture complex customer behavior and their sensitivity to model misspecification. This research proposes a novel, model-agnostic framework that leverages guided discrete diffusion to overcome these limitations. The new approach represents product assortments as binary vectors and employs a learned reverse diffusion process for stochastic search, thereby avoiding the need for explicit combinatorial enumeration. To integrate decision objectives, a reward-guided mechanism is introduced. This mechanism biases local transitions based on estimates of expected revenue, effectively balancing the exploration of new assortment configurations with the exploitation of promising ones during the generation process. Empirical results demonstrate that this diffusion-based method consistently identifies high-quality assortments. It exhibits robustness even when underlying customer choice models are misspecified, frequently recovering near-optimal solutions in complex, high-dimensional scenarios. Furthermore, the generative nature of diffusion allows for the production of diverse high-performing assortments, offering greater flexibility than a single deterministic solution. These findings highlight the significant potential of generative modeling as a scalable and robust paradigm for data-driven combinatorial optimization.