Abstract
Local life service is a vital scenario in Kuaishou App, where videorecommendation is intrinsically linked with store's location information. Thus,recommendation in our scenario is challenging because we should take intoaccount user's interest and real-time location at the same time. In the face ofsuch complex scenarios, end-to-end generative recommendation has emerged as anew paradigm, such as OneRec in the short video scenario, OneSug in the searchscenario, and EGA in the advertising scenario. However, in local life service,an end-to-end generative recommendation model has not yet been developed asthere are some key challenges to be solved. The first challenge is how to makefull use of geographic information. The second challenge is how to balancemultiple objectives, including user interests, the distance between user andstores, and some other business objectives. To address the challenges, wepropose OneLoc. Specifically, we leverage geographic information from differentperspectives: (1) geo-aware semantic ID incorporates both video and geographicinformation for tokenization, (2) geo-aware self-attention in the encoderleverages both video location similarity and user's real-time location, and (3)neighbor-aware prompt captures rich context information surrounding users forgeneration. To balance multiple objectives, we use reinforcement learning andpropose two reward functions, i.e., geographic reward and GMV reward. With theabove design, OneLoc achieves outstanding offline and online performance. Infact, OneLoc has been deployed in local life service of Kuaishou App. It serves400 million active users daily, achieving 21.016% and 17.891% improvements interms of gross merchandise value (GMV) and orders numbers.