TLDR: OneLoc is a new AI-powered recommendation system for Kuaishou’s local life service that uses a generative approach to suggest videos based on user interests and real-time location. It integrates geographic information deeply through “geo-aware semantic IDs,” “geo-aware self-attention,” and “neighbor-aware prompts.” It also balances user preferences with business goals (like sales) using reinforcement learning with geographic and GMV rewards. Deployed on Kuaishou, OneLoc serves 400 million daily users and has significantly boosted GMV and order numbers.
The Kuaishou App, a major internet platform, features a vital “local life service” where video recommendations are closely tied to a store’s location. This presents a unique challenge: how to recommend videos that not only match a user’s interests but also consider their real-time location. For instance, two users in the same spot might both want a drink. If one user frequently interacts with KFC videos, but McDonald’s is closer, the system might recommend McDonald’s. Conversely, if another user often buys from Starbucks, the system might recommend Starbucks even if it’s further away, prioritizing their strong preference. This highlights the importance of balancing user interest with geographical proximity.
To tackle this, researchers have introduced OneLoc, an innovative end-to-end generative recommendation model specifically designed for local life services. This new approach moves away from traditional recommendation methods, embracing a generative paradigm that has shown success in other areas like short video recommendations (OneRec), search suggestions (OneSug), and advertising (EGA).
OneLoc addresses two primary challenges. The first is effectively utilizing geographic information. Previous methods often treated location as a separate feature or used it only for guiding the recommendation output. OneLoc takes a more comprehensive approach by integrating geographic data at multiple levels. It uses “geo-aware semantic IDs” which combine video and location information for better categorization. It also employs “geo-aware self-attention” within its processing unit, considering both video location similarity and the user’s current location. Furthermore, a “neighbor-aware prompt” captures rich contextual information from the user’s surroundings to guide the recommendation generation.
The second challenge OneLoc addresses is balancing multiple objectives, such as user interests, the distance to stores, and other business goals like increasing sales. To achieve this, OneLoc uses a technique called reinforcement learning. It introduces two specific “reward functions”: a geographic reward that gives higher scores to closer locations, and a GMV (Gross Merchandise Value) reward that encourages recommendations leading to more purchases. By optimizing for these rewards, OneLoc can fine-tune its recommendations to meet both user needs and business targets.
The effectiveness of OneLoc has been demonstrated through extensive testing. In offline experiments using Kuaishou’s large dataset, OneLoc significantly outperformed existing recommendation models, showing improvements in metrics like Recall and NDCG. For example, on the KuaiLLSR dataset, it achieved a 13.46% increase in Recall@5 and a 14.47% increase in NDCG@5.
Crucially, OneLoc has been successfully deployed in the Kuaishou App’s local life service. It currently serves 400 million active users daily. Online A/B tests showed remarkable improvements in key business metrics: a 21.016% increase in Gross Merchandise Value (GMV) and a 17.891% increase in the number of orders. This real-world deployment confirms OneLoc’s ability to deliver superior performance in a high-traffic industrial environment.
The system’s architecture involves a Trainer for offline learning, an Inference Server for real-time recommendations, a Video Mapping Server to link recommendations to actual videos, and a Reward System for scoring. The deployment also incorporates performance optimizations like mixed-precision computation and TensorRT acceleration to handle the massive user base efficiently.
Also Read:
- TBGRecall: Enhancing E-commerce Product Recommendations with Next Session Prediction
- New Approach to Recommender Systems: L3AE Combines LLMs and Linear Models
This research marks a significant step forward in location-based recommendation systems, particularly for local life services, by effectively integrating geographic awareness and multi-objective optimization into a generative framework. For more technical details, you can refer to the full research paper available here.
