TAMOR: Tier-Aware Multi-Objective Recommendation for Ant Fortune Financial Marketing

Abstract

Online marketing recommendation is crucially important for user growth of mobile applications. However, there are currently three common challenges in designing such an efficient recommendation system. First, on the user side, users can be stratified into different layers which have distinctive user characteristics and marketing objectives. Second, on the item side, items from heterogeneous business scenarios need to be mixed together for ranking. Third, there are often multiple marketing objectives, which are even internally related to each other. In this paper, we address the above challenges by proposing a joint training system T ier- A ware M ulti- O bjective R ecommendation ( TAMOR ). The TAMOR system leverages all tiers of data to train a unified model, while the representation learned by the model for users and items are aware of data tiers. Besides, in order to better deal with the multi-objective prediction problem, the user bias learning is designed to learn user preferences, which are then used to assist learning for user-specific tasks. TAMOR has been deployed for financial marketing of Ant Fortune, which brings a 10.67% boost for the number of daily new high-holding users.