UCL School of Management is delighted to welcome Yicheng Song, Questrom School of Busines, to host a research seminar discussing ‘When diversity becomes relevant - A multi-category ultility model of consumer response to content recommendations.’
The diversity of a set of recommendations can improve consumers’ satisfaction with the personalized recommender system. However, diversifying a list of products for a one-shot recommendation sacrifices relevance, which can reduce its value. We identify a popular scenario, sessions of online news consumption, where one can increase the diversity of recommendations over the entire session while improving the relevance of each recommendation within the session. Our approach is based on a multi-category utility model that captures consumers’ preference towards different types of content, how quickly they satiate with one type and substitute it with another, and how they trade off their own costly search efforts with selecting from the recommended articles to find new content. Taken together, these three elements enable us to characterize how utility maximizing consumers construct diverse “baskets” of content over the course of each session, and how likely they are to click on content recommended to them. We estimate this model using a clickstream dataset from a large international media outlet and apply it to determine the most relevant content at different stages of an online session. We demonstrate that by taking into account how consumers sequentially select content from different categories over time within a session, we not only recommend more diverse content over a session, but also recommend more relevant content than methods that do not incorporate such information. The diversity of the content recommended by the proposed approach closely matches the diversity sought by individual readers in their actual natural consumption—exhibiting the lowest concentration-diversification bias when compared to other personalized recommender systems. Meanwhile, the proposed approach makes 6%–14% more accurate recommendations than optimized alternatives. Using a policy simulation, we estimate that recommending content using the proposed approach would result in visitors reading 57% additional articles at the studied website, which has direct revenue implication for the publisher of this site.