Revenue-based personalized product recommendation considering stochastic purchase probability

Recommender systems(RS) play a critical role in e-commerce platforms by providing personalized and relevant product suggestions to customers, thereby enhancing their shopping experience and increasing platform revenue. Existing RSs focus on improving accuracy or maximizing user purchase probability when generating recommendations. However, a sole emphasis on accuracy does not ensure the optimization of platform revenue, and recommendations that maximize user purchase probability can also fail to simulate the real purchase behavior of users, which shows strong uncertainty due to external factors. To address these issues, we propose a two-stage personalized product recommendation method based on stochastic purchase probability (PRSPP). In the first stage, we follow previous studies which prioritize user preferences during the recommendation process. A taxonomy-based approach is employed to estimate user preferences at the category level and select candidate products for each user. Subsequently, considering the impact of factors such as product price, sales and category similarity on user utility, a logistic regression model is employed to quantify user preferences for these candidate products and further estimate user purchase probability for them. In the second stage, we aim to optimize the recommendation from the perspective of platform operators, considering user purchases are subject to diverse external factors and exhibit strong uncertainty. We treat purchase probability as a random variable, and a stochastic optimization model with the objective of maximizing platform revenue is formulated. Furthermore, we apply the Sample Average Approximation (SAA) approach to solve the model. Finally, we conduct experiments on Amazon public dataset, and the results present advantages of PRSPP in improving both recommendation accuracy and platform revenue.