Dynamic Forgetting and Spatio-Temporal Periodic Interest Modeling for Local-Life Service Recommendation

TL;DR

This work tackles the challenge of modeling sparse, long user behavior sequences in local-life services by introducing STIM, a spatiotemporal periodic interest model. STIM leverages a forgetting-curve–driven dynamic masking to emphasize recent and periodic behaviors, a query-based mixture of experts to adaptively fuse time, location, and item signals, and a hierarchical multi-interest network to capture diverse user intents. The approach yields strong offline gains and a notable online improvement in Gross Transaction Volume, while being deployed in large-scale local-life systems. By explicitly modeling recency, periodicity, and spatiotemporal interactions, STIM provides a robust framework for long-sequence recommendation under sparsity and strong context dependence, with practical impact on e-commerce and local-service platforms.

Abstract

In the context of the booming digital economy, recommendation systems, as a key link connecting users and numerous services, face challenges in modeling user behavior sequences on local-life service platforms, including the sparsity of long sequences and strong spatio-temporal dependence. Such challenges can be addressed by drawing an analogy to the forgetting process in human memory. This is because users' responses to recommended content follow the recency effect and the cyclicality of memory. By exploring this, this paper introduces the forgetting curve and proposes Spatio-Temporal periodic Interest Modeling (STIM) with long sequences for local-life service recommendation. STIM integrates three key components: a dynamic masking module based on the forgetting curve, which is used to extract both recent spatiotemporal features and periodic spatiotemporal features; a query-based mixture of experts (MoE) approach that can adaptively activate expert networks under different dynamic masks, enabling the collaborative modeling of time, location, and items; and a hierarchical multi-interest network unit, which captures multi-interest representations by modeling the hierarchical interactions between the shallow and deep semantics of users' recent behaviors. By introducing the STIM method, we conducted online A/B tests and achieved a 1.54\% improvement in gross transaction volume (GTV). In addition, extended offline experiments also showed improvements. STIM has been deployed in a large-scale local-life service recommendation system, serving hundreds of millions of daily active users in core application scenarios.

Figures (10)

• Figure 1: Ebbinghaus's forgetting curve in memory and recommendation scenarios.
• Figure 2: Overall architecture of the spatio-temporal periodic interest model.
• Figure 3: Illustration of forgetting curve decay in recommendation scenarios. The rightmost block depicts the request query, the white blocks indicate padding, and the left blocks represent historical user behavior sequences ordered chronologically from the farthest to the nearest.
• Figure 4: Comparison of the foundational functions of forgetting curves.
• Figure 5: Effects of hyperparameters on the forgetting curve of TRec.