Random Utility Models with Cardinality Context Effects for Online Recommendations

Abstract

This work creates new random utility models able to capture a cognitive heuristic known as consideration set theory. These choice models, which capture cardinality context effects in addition to item and user attributes, are ideal for recommending personalized assortments in online subscription service platforms. In such situations, users may be overwhelmed by a large amount of options, increasing their no-choice probability. Our new models modify the multinomial logit model by subtracting from an item’s perceived representative utility a disutility caused by the cardinality of the recommended assortment. Because this disutility does not affect the no-choice option, the user’s net benefit from evaluating a certain number of items is captured as a trade-off between the benefits and the costs of considering those items. Using an empirical data set capturing changing no-choice probabilities for different assortment sizes, a gradient descent algorithm estimates model input parameters. An algorithm is presented to efficiently solve the optimal assortment problem under cardinality context effects with equal profit margin. Computational experiments and statistical tests using the MovieLens dataset analyze the sensitivity of model parameters on the optimal assortment cardinality and no-choice probability.