This paper proposes OneLoc, a novel end-to-end generative recommendation model for the Kuaishou app's local lifestyle service recommendation system. Unlike existing end-to-end models that only consider user interests, OneLoc simultaneously considers user interests and real-time location information to achieve recommendations. To achieve this, we propose three techniques (geo-aware semantic ID, geo-aware self-attention, and neighbor-aware prompt) that leverage geographic information from various perspectives, as well as two reinforcement learning-based reward functions (geographic reward and GMV reward). OneLoc has been deployed on the Kuaishou app, serving 400 million active users, and has achieved performance improvements of 21.016% in GMV and 17.891% in the number of orders, respectively.
