58 / 2024-04-18 09:11:30

An End-to-end Model for Share-of-choice Product Line Design Problem

Product Line Design,Prescriptive Analytics,Polyhedral Estimation,Adaptive Conjoint Analysis.

This paper studies the share-of-choice product line design problem, where decision-makers aim to maximize

the market share, i.e., the percentage of the consumers to whom at least one product delivers a non-negative

utility. Because the customers’ utilities are not directly observable, the input utilities of the existing models

need to be estimated from some primitive data. The conjoint analysis, which asks respondents to evaluate

a set of products, is the main source of such data. However, with the limited survey questions, the utilities

cannot be precisely estimated, and therefore, the models that plugin the estimated utilities cannot identify

the optimal solution and corresponding market share. In this study, by applying the polyhedral method in

utility calibration and robust optimization techniques, we propose a novel end-to-end model inputting the

survey data and outputting the data-driven design. Importantly, we provide guarantees of the market share

of the returned product line design. For a general product line design problem, we provide the guarantees

related to the size of the set containing all possible utilities consistent with the survey data. For a single

product design special case, our model can identify the optimal design within an explored set of all the

linear combinations of the delivered questions and guarantee the market share even if the uncertainty sets

are unbounded. We further develop an adaptive conjoint analysis where new questions better align with

the proposed model are added based on its solution. The proposed adaptive conjoint analysis is proven to

recover the customer utilities within finite rounds. Through extensive numerical experiments, we show that

the proposed model significantly outperforms the “estimate-then-optimize” model and the proposed adaptive

conjoint analysis can generate more efficient questions to fit the proposed model.