Integrating Empirical Estimation and Assortment Personalization for E-Commerce: A Consider-Then-Choose Model

Abstract

We develop a new approach that integrates empirical estimation and assortment optimization to achieve display personalization for e-commerce platforms. We propose a two-stage Multinomial Logit (MNL) based consider-then-choose model, which accurately captures the two stages of a consumer's decision-making process -- consideration set formation and purchase decision given a consideration set. To calibrate our model, we develop an empirical estimation method using views and sales data at the aggregate level. The accurate predictions of both view counts and sales numbers provide a solid basis for our assortment optimization. To maximize the expected revenue, we compute the optimal target assortment set based on each consumer’s taste. Then we adjust the display of items to induce this consumer to form her consideration set that coincides with the target assortment set. We formulate this consideration set induction process as a nonconvex optimization, for which we provide the sufficient and necessary condition for feasibility. This condition reveals that a consumer is willing to consider at most K(C) items given the viewing cost C incurred by considering and evaluating an item, which is intrinsic to consumers’ online shopping behavior. As such, we argue that the assortment capacity should not be imposed by the platform, but rather comes from the consumers due to limited time and cognitive capacity. We provide a simple closed-form relationship between the viewing cost and the number of items a consumer is willing to consider. To mitigate computational difficulties associated with nonconvexity, we develop an efficient heuristic to induce the optimal consideration set. We test the heuristic and show that it yields near-optimal solutions. Given accurate taste information, our approach can increase the revenue by up to 35%. Under noisy predictions of consumer taste, the revenue can still be increased by 1% to 2%. Our approach does not require a designated space within a webpage, and can be applied to virtually all webpages thereby generating site-wise revenue improvement.