Spacetime-GR: Amap's Spatiotemporal Generative Model for Industrial POI Recommendation

ICDM 2026 Overview

ICDM 2026 (IEEE International Conference on Data Mining) received 1,966 submissions, with 516 in the Applied track and 103 papers accepted (19.96% acceptance rate). ICDM is one of the three top-tier international conferences in data mining alongside KDD and SDM, and is a CCF Class B recommended conference.

Introduction

In map applications and local-life platforms, users frequently ask "Where should I go next?" This seemingly simple question involves a complex spatiotemporal recommendation task. User interests vary significantly by time of day, location, and travel scenario: lunch hours favor restaurants, afternoons favor cafés; local commuting emphasizes daily services, while unfamiliar cities shift interest to attractions, hotels, and transport hubs.

Traditional recommendation models excel at static items but struggle with POIs that heavily depend on spatiotemporal context. Industrial scenarios add massive scale (hundreds of millions of POIs, massive users, real-time online requests), demanding models that are not only accurate but also fast and deployable.

Against this background, the Amap team proposed Spacetime-GR : a spatiotemporal-aware generative recommendation model. Its goal is not to replace existing systems but to harness the unified modeling power of generative models for POI scenarios through purpose-built encoding and training pipelines, achieving: (1) sensitive understanding of spatiotemporal context; (2) efficient modeling of ultra-large POI vocabularies; (3) practical adaptation to online recommendation systems. Spacetime-GR demonstrates that generative recommendation can run in real industrial systems and deliver measurable gains.

Paper link:

https://arxiv.org/abs/2508.16126

Project Introduction

Spacetime-GR is a complete generative framework for industrial POI recommendation, comprising three key components:

Hierarchical POI Encoding

The biggest engineering challenge in POI recommendation is vocabulary size. Treating each POI as an independent token is infeasible at hundreds of millions of locations. Spacetime-GR adopts a geographic-aware hierarchical POI indexing strategy : first, partition POIs into geographic blocks; second, assign an inner ID within each block. A POI is uniquely represented by a (block, inner) token pair. This yields two direct benefits: (1) drastically reduces vocabulary size, making softmax computation feasible; (2) naturally injects geographic information because the encoding itself embeds spatial structure. The model reasons in a hierarchy — first selecting a region, then a specific point — aligning with real geographic distributions.

Spatiotemporal Encoding Module

Unlike traditional sequence recommendation, POI recommendation is driven by both time and space. Spacetime-GR designs a dedicated spatiotemporal encoding module that unifies the following signals as tokens in the sequence: time (month, day-of-week, day, hour); user geographic location; POI geographic location; POI category; behavior type (e.g., click or functional action); user profile information. Unlike methods that treat spatiotemporal features as side inputs, Spacetime-GR feeds them directly as sequence tokens, allowing attention layers to naturally capture: (1) user preference shifts in specific time windows; (2) interest differences between local and travel scenarios; (3) associations between POI semantics and geographic neighborhoods.

Multi-Stage Training and Deployment

To serve industrial systems, Spacetime-GR employs a practical three-stage pipeline: (1) Pre-training on massive user behavior sequences to learn general patterns; (2) SFT (Supervised Fine-Tuning) for downstream recommendation tasks; (3) Alignment (DPO) to further align with user click preferences, enhancing end-to-end recommendation capability. Beyond generating the next POI, the model outputs user/POI embeddings, ranking scores, and end-to-end recommendation candidates, enabling flexible integration into existing architectures without a full rebuild.

Research Background

Industry Demand: Map Recommendation as High-Frequency Necessity

With the rise of local-life services, map products have evolved from navigation tools into intelligent platforms integrating search, discovery, recommendation, and decision-making. Users decide not just how to reach a place but where to go — dining, leisure, shopping, travel, work — all generating POI recommendation demand. The system must understand short-term intent and long-term preferences, balance geographic accessibility and temporal context, and optimize both click-through and business conversion.

Core Challenge: Spatiotemporal Sensitivity + Ultra-Large Scale

POI recommendation differs fundamentally from traditional item recommendation because user preferences are extremely sensitive to spatiotemporal conditions. The same user exhibits completely different behavioral intents across locations and times. Industrial POI scales reach tens to hundreds of millions, causing: huge candidate space; difficulty training on long-tail POIs; prohibitive cost of direct large-model generation.

Limitations of Existing Methods

Current generative recommendation approaches in industrial POI settings suffer from vocabulary explosion, insufficient spatiotemporal modeling, and inference efficiency constraints. A truly industrial-grade generative, spatiotemporal-aware, deployable model is therefore critical.

Highlights

Spacetime-GR's contributions span model design, training strategy, and deployment capability:

Spatiotemporal-aware generative framework for online POI recommendation: A complete solution from task definition, not a simple retrofit. It targets online "click recommendation" rather than offline check-in prediction, aligning with real product needs.

Hierarchical encoding solves large-vocabulary challenge: The block+inner two-level POI encoding compresses the massive POI vocabulary to a trainable scale, reducing model complexity and enhancing geographic structure modeling.

Spatiotemporal context fused into main sequence modeling: Instead of treating time and location as auxiliary features, they are input as sequence tokens, enabling attention layers to learn contextual dependencies and improving spatiotemporal sensitivity.

Supports "representation learning + ranking + end-to-end recommendation" triple mode: Through pre-training, SFT, and DPO alignment, Spacetime-GR serves different industrial modules.

Real online deployment achieved: The work goes beyond paper metrics, validated in industrial recommendation systems with clear CTR and CVR improvements.

Experimental Results

SFT Stage: Significant Online Ranking Improvement

On industrial data, both Embedding-based ranking SFT and Generative ranking SFT were tested. Both boosted the existing online ranking system, with Generative ranking SFT performing better, indicating deep interaction modeling is crucial for POI recommendation. Combining both SFT outputs raised AUC from 0.7043 to 0.7385 , showing the model effectively supplements information the online system cannot directly capture.

Alignment Stage: Enhanced Joint Retrieval and Ranking

After DPO alignment, Spacetime-GR's recommendation quality further improved. Both LLM-based automatic evaluation and human evaluation showed the aligned model significantly outperformed the original online system, confirming the model not only predicts likely clicks but also better matches true interest distributions.

Online Experiments: Clear Business Metric Gains

In live traffic experiments, Spacetime-GR delivered tangible business value: in Scenario A, CTR increased ~6% and CVR increased ~4.2% ; in Scenario B, PV CTR increased ~15.4% and UV CTR increased ~7.1% . This demonstrates practical business impact beyond academic improvements.

Public Dataset Validation: Strong Generalization

On public POI datasets Foursquare-NYC, Foursquare-TKY, and Gowalla-CA, Spacetime-GR showed strong competitiveness, achieving state-of-the-art results on several datasets, indicating the method transfers well beyond industrial online settings.

Conclusion

Spacetime-GR advances generative recommendation from "feasible" to "usable, effective, and deployable." Through hierarchical POI representation, explicit spatiotemporal encoding, multi-modal enhancement, and multi-stage training, the model resolves core contradictions in large-scale online POI recommendation: large vocabulary vs. high efficiency; spatiotemporal sensitivity vs. unified modeling; academic performance vs. industrial deployment. This work has already delivered real business gains, proving the immense potential of generative models in map recommendation and local-life services.