Spatiotemporal-Enhanced Network for Click-Through Rate Prediction in Location‑Based Services

The paper introduces a novel recommendation model called StEN (Spatiotemporal‑Enhanced Network) designed for click‑through rate (CTR) prediction in location‑based services (LBS) such as food delivery. It argues that user behavior in LBS is strongly tied to both time and space, and that existing industrial recommendation pipelines only perform coarse LBS recall without fine‑grained spatiotemporal modeling.

Related Work – The authors review several recent works that incorporate temporal or spatial signals, including Regularized Adversarial Sampling and Deep Time‑aware Attention (CIKM 2019), Calendar Graph Neural Networks (KDD 2020), Deep Time‑Aware Item Evolution Network (CIKM 2020), TLSAN (2021), TiSASRec (WSDM 2020), and TimelyRec (WWW 2021). These works either focus on temporal patterns, spatial features, or both, but none jointly model static spatiotemporal characteristics, dynamic preferences, and target‑item attention as proposed here.

Model Overview – StEN consists of three major modules:

Static Spatiotemporal Feature Activation (StPro) : extracts inherent spatiotemporal attributes from relatively static features (user, item, context, target query) such as hour, weekday, geohash, AOI, and activates them via a DIN‑style gating mechanism.

Dynamic Spatiotemporal Preference Activation (StPre) : contains three sub‑modules – (a) Temporal Order Activation (TEA) that computes time‑difference weights with exponential smoothing and softmax normalization; (b) Temporal Periodic Fusion (TPF) that splits user clicks into time‑type buckets (breakfast, lunch, etc.) and averages their embeddings; (c) Spatial Preference Activation (SPA) that uses geohash/AOI features to generate a sigmoid gating factor for dynamic click sequences.

Spatiotemporal Target Attention (StTA) : augments the classic target‑item attention by generating adaptive query/key/value projection matrices conditioned on spatiotemporal features, enabling the attention mechanism to capture time‑space dependencies between the target item and historical clicks.

The outputs of StPro, StPre, and StTA are concatenated (along with a simple mean‑pooling branch) and fed into a DNN trained with binary cross‑entropy loss.

Data Investigation – The authors analyze user demand variations across different times of day and location types, illustrating the necessity of fine‑grained spatiotemporal modeling for food‑delivery scenarios.

Experiments – Offline evaluations are conducted on two proprietary industrial datasets (store‑level and product‑level) derived from a major food‑delivery platform. StEN achieves state‑of‑the‑art performance compared with four baseline methods, and ablation studies confirm the contribution of each module. Online A/B testing shows a 1.6% lift in CTR and a 2.1% increase in orders after deploying StEN.

Conclusion & Outlook – StEN effectively captures static spatiotemporal traits, dynamic preferences, and target‑aware attention, delivering superior CTR prediction in LBS. Future work will iterate on the model to further enhance its industrial applicability.