TLDR: A new AI model called STIM improves local-life service recommendations by mimicking how human memory works, addressing challenges like sparse user data and location/time-based preferences. It uses a ‘forging curve’ to prioritize recent and periodic interests, a ‘mixture of experts’ to adapt to different scenarios, and a ‘multi-interest network’ to capture diverse user preferences, leading to significant improvements in transaction volume and performance in real-world applications.
In today’s fast-paced digital world, recommendation systems are crucial for connecting users with countless services, especially in the realm of local-life platforms like food delivery or shopping. However, these systems face significant hurdles: user behavior data can be sparse, meaning there are gaps in activity, and user preferences are heavily influenced by when and where they are. Imagine trying to recommend a restaurant to someone whose last interaction was a year ago, or understanding their daily coffee habit versus a once-a-month special dinner. This complexity makes it hard for traditional systems to accurately predict what users want.
Drawing inspiration from how human memory works, researchers have developed a novel approach called Spatio-Temporal Periodic Interest Modeling, or STIM. This method addresses the unique challenges of local-life service recommendations by mimicking the natural processes of forgetting and remembering, particularly the ‘recency effect’ and the cyclical nature of memory.
How STIM Works: A Memory-Inspired Approach
STIM integrates three core components to understand and predict user interests more effectively:
First, the **Dynamic Masking Module** is based on the concept of the ‘forgetting curve.’ Just as our memory of an event fades over time but is reinforced by repeated exposure, STIM uses this principle to weigh user behaviors. It groups user activities by time (hourly, weekly) and location (using ‘geohash’ codes) and treats recent or recurring activities within these groups as ‘review points.’ This allows the system to prioritize the most relevant recent behaviors while also capturing long-term periodic habits, effectively filtering out less important, older data.
Second, the **Query-Based Mixture of Experts (MoE)** module acts like a team of specialized consultants. When a user makes a query (e.g., searching for a restaurant), this module adaptively activates different ‘expert’ networks—one for time, one for location, and one for items. This allows the system to deeply understand the combined influence of these factors. For instance, it can recognize that a user’s interest in coffee shops might be strong on weekday mornings but shifts to dinner restaurants on weekend evenings. It even incorporates special factors like holidays, adjusting recommendations accordingly.
Third, the **Hierarchical Multi-Interest Network Unit (HMIN-Unit)** is designed to uncover the diverse interests a user might have. It processes user behavior data in two stages: a ‘shallow interaction’ to identify basic connections between what a user is looking for and their past actions, and a ‘deep interaction’ to capture more complex, context-sensitive preferences. This helps the system build a rich profile of a user’s multiple interests, from their daily routines to their occasional indulgences.
Also Read:
- Navigating Urban Futures: A New AI Model Predicts Next Steps on City Roads
- Brain-Inspired AI Helps Self-Driving Cars Remember Past Experiences
Real-World Impact and Performance
The STIM method has been rigorously tested and deployed in a large-scale local-life service recommendation system, serving hundreds of millions of daily active users. Online A/B tests showed a significant 1.54% improvement in Gross Transaction Volume (GTV), demonstrating its substantial business value. Offline experiments also confirmed its superior performance across various metrics like Click-Through Rate (CTR) and Click-to-Conversion Rate (CTCVR) on large datasets like Ele.me and TRec.
Notably, STIM excels in ‘cold start’ scenarios, where user behavior sequences are very sparse (e.g., new users or users with very infrequent activity). Its dynamic masking effectively handles these situations by focusing on the most recent available behaviors, proving its robustness even with limited data.
In essence, by intelligently mimicking the nuances of human memory, STIM provides a powerful and adaptive solution for personalizing recommendations in the complex and dynamic world of local-life services. For more technical details, you can refer to the full research paper here.
