Context dependent discrete choice models and assortment optimization for online retail

Abstract

Substantive research in the cognitive, psychological, and user research fields has identified that high assortment cardinality can lead to ‘information overload’-the phenomenon in which choice deferral increases when the number of items recommended is high. We develop post-filtering descriptive models to capture individual user behavior under information overload effects, in which choice deferral increases with assortment cardinality. By defining the observable part of item utility (called representative utility) as a function of assortment size, in addition to the attributes of each item and those of the user, a new family of random utility models are created. The RUMs devised in this work capture: 1) Trade-offs between information overload and overall assortment utility (All four models); 2) the notion of novelty and diversity by recommending items that the user would not find herself (two non-linear models); and 3) the moderating influence of the viewing medium (phone vs. webpage) causing presentation bias (One linear and one non-linear model).

The primary objective of this research is to develop descriptive mathematical models capturing context-effects associated with individual user selection behavior found in marketing and behavioral research. These models are then used as inputs to assortment optimization problems to optimize personalized recommendations in an online retail environment. Our work makes contributions at the intersection of three fields: (1) discrete choice models (2) recommender systems and (3) e-commerce assortment planning.

These four models are extensions to the Multinomial Logit model (MNL) and are termed the MNL-CE family. The linear extension for the Nested Logit Model (NLM) is also formulated as NLM-CE. Properties of the extended models are proved and compared to existing models. The developed models are used as inputs to assortment optimization problem to determine the optimal assortment of items to offer to each user under two scenarios: 1) All items available have comparable profits 2) All items available have unequal profits. The MNL-CE models are analyzed for the structure of the optimal assortment under cardinality context-effects and efficient algorithms are developed and tested on two real world datasets (MovieLens, UCI Repository Online Retail dataset). A comprehensive statistical analysis reveals the significant impact of MNL-CE model parameters on the optimal solution. For the NLM-CE, the assortment optimization model is shown to be NP-Hard.