UC Santa Cruz

Product recommender systems have become increasingly important, as consumers are exposed to massive amounts of product information on the Internet and product recom- mender systems can suggest products that interest consumers. A lot of work has been done over the past few decades to develop recommendation algorithms. Most existing algorithms reduce the recommendation problem to rating prediction and measure the recommendation quality by Root Mean Squared Error (RMSE). The existing algorithms have proven successful in a wide range of applications including movie , news and music recommendations.

However, when it comes to product recommendation, there are specific aspects the recommendation algorithm needs to consider. First, price is important and the rec- ommendation algorithm needs to consider consumer’s Willingness-to-Pay (WTP); sec- ond, it is important to consider inter-product relationships; third, it is necessary to take account of the benefits to both consumer and producer. Without such consideration, recommendation algorithms might perform poorly. For example, the recommendations are too expensive for the consumer to buy, or the recommendations do not complement the products the consumer already has, or the system prices the products in favor of the consumers with little consideration of the benefit to producers. Unfortunately, these

three aspects are not well addressed in existing recommendation algorithms and pose challenges to product recommendation.

In this dissertation, we propose to address these aspects by leveraging well- established economic principles. In particular, we adopt producer surplus, consumer sur- plus and total surplus to represent the benefits of producer, consumer and the platform, respectively. For producers, we propose to elicit consumer WTP in the E-commerce setting and perform personalized promotion based on the estimated WTP. The goal of personalized promotion is to maximize producer’s profit. For consumers, we propose to recommend by multi-product utility maximization. Our proposed method can auto- matically learn the relationship from real-world transaction data. For the platform, we propose and implement a Total Surplus Maximization (TSM) based recommendation framework, in which the benefits of both producer and consumer are considered. The TSM framework can conveniently specialize into several typical applications, including E-commerce, P2P lending and the online freelancer market. The proposed methods have been evaluated in real-world datasets and the experimental results demonstrate that the proposed methods are advantageous to existing algorithms.