Pan Li, an Information System PhD candidate at NYS Stern, studies the application of personalization techniques in business, where he aims to improve the design of recommendation models that effectively explore consumer preferences and improve consumer satisfaction.
Modeling consumer preference is an important research topic for the design of recommender systems. However, it is not a trivial task as consumer preferences are typically multi-faceted during their online purchase sessions, and they may have various types of intentions when interacting with recommendation agents, such as locating the most useful products or exploring novel content. Although the current generation of recommender systems has achieved great success in industrial applications, they primarily focus on only one dimension of consumer preference, such as identifying the most similar candidate products for the target consumer, while not addressing consumers’ multi-faceted preferences of product variety, novelty, freshness, and so on. As a result, they would only reach sub-optimal recommendation performance in business practice and might not fully meet consumers' satisfaction.
In my research, I have introduced and developed two research streams of novel recommender system methods, namely the unexpected recommender system and cross-domain recommender system, to model consumer desire of seeking novel content in recommendations. In particular, the former focuses on "within-domain" consumer preference exploration, as it aims at providing novel and useful recommendations simultaneously from the same product domain to alleviate the consumer boredom problem. Meanwhile, the latter learns consumer preference information across multiple product domains and then produces better recommendations across these different domains to expand the consumers' horizons. Both methods would be effective and efficient in exploring consumer interests and improving consumer online experience, as I empirically demonstrate through an extensive set of simulation, offline, and online experiments on multiple industrial recommendation applications. Moreover, to further demonstrate the economic impact of my proposed approach and to provide managerial implications for practitioners, I also conduct several large-scale randomized online controlled experiments at Alibaba, which illustrate significant business performance improvements over the latest production system, leading the company to move the proposed models into deployment.
