Spacetime‑GR: AI‑Powered Spatiotemporal Model Transforming POI Recommendations

Looking back, Amap has served over one billion users as a leading map app. Looking forward, it aims to combine two decades of data with modern large‑model capabilities, turning the map from a static base‑map + passive planning system into a dynamic, cognition‑aware, decision‑making platform.

Innovation Recommendation Team – Home Recommendation Group: G‑where predicts likely destinations, G‑action forecasts travel needs in real‑time contexts, and G‑plan uses a large model to assemble fragmented user intents into complete schedules.

Innovation Recommendation Team – Human‑Place Model Group: Pre‑trains a massive spatiotemporal model on anonymized app behavior data to learn city‑level rhythms, then fine‑tunes it into a domain expert that delivers precise POI suggestions.

Innovation Recommendation Team – Exploration Recommendation Group: Leverages world knowledge to aggregate user interests and surface surprising content themes.

Project Background

Generative recommendation, typically built on large language models (LLMs), has become the dominant paradigm, outperforming traditional cascade systems (recall → coarse‑ranking → fine‑ranking). In Amap’s POI recommendation scenario, the goal is to predict locations a user may like based on historical preferences and current spatiotemporal context. Real‑world deployment faces three major challenges:

Massive vocabularies: Modeling hundreds of millions of POIs leads to an infeasible token space.

Spatiotemporal sensitivity: User behavior sequences contain both POI attributes and time‑location context, causing interest shifts across different times and places.

Encoding low‑frequency POIs: Long‑tail POIs appear rarely, making them hard to represent effectively.

To address these challenges, the Innovation Recommendation Team proposes Spacetime‑GR , a spatiotemporal‑aware generative recommendation model for large‑scale online POI recommendation. The model modifies the standard LLM framework by introducing a hierarchical POI index and directly encoding spatiotemporal information as tokens, thereby improving both computational efficiency and recommendation quality.

Model Architecture

The hierarchical POI index uses two tokens per POI: a block token representing the geographic block and an inner token for the POI’s position within that block. This reduces the vocabulary from hundreds of millions to around 400 k tokens and enables the model to first predict the most relevant block before selecting a POI inside it.

Spatiotemporal encoding adds three additional tokens to each user action: u_i (user identifier), block_i, inner_i, and a_i (action type such as click, navigation, ride‑hailing). Side‑information like POI category and precise location is fused into the inner_i representation. The loss function is a cross‑entropy that only back‑propagates through tokens corresponding to interest‑type behaviors, ensuring the model focuses on genuine user preferences.

Training Stages

Stage 1 – Global behavior pre‑training: The model learns latent patterns by predicting the next user action across the entire Amap behavior corpus.

Stage 2 – Downstream fine‑tuning: Business‑specific data (search promotion, etc.) fine‑tunes the model for particular recommendation strategies.

Data Processing

Data cleaning occurs at two levels. At the behavior level, actions are classified as functional (e.g., navigation to a known place) or interest‑driven (e.g., browsing entertainment POIs). Only interest‑driven actions are used for loss calculation. At the sequence level, a “richness” metric filters out low‑value sequences that contain many repeated POIs, retaining diverse user journeys for training.

Curriculum Learning Strategy

Training data are split into single‑mode (one spatiotemporal state per sub‑sequence: local, pre‑trip, in‑trip) and multi‑mode datasets. The model first learns on the simpler single‑mode data and then gradually incorporates the more complex multi‑mode data, improving its ability to predict behavior transitions.

Experimental Results

Pre‑training effect: Using hit‑rate (HR@k) as the evaluation metric, Spacetime‑GR consistently outperforms baseline cascade models.

Scaling law: Larger parameter counts yield higher pre‑training performance, mirroring trends observed in generic LLMs.

Discovery rate: A new metric D(k,m) measures the novelty of recommended POIs (those not interacted with in the past m days). Spacetime‑GR achieves higher discovery rates across all settings, indicating more innovative recommendations.

Conclusion

The work presents Spacetime‑GR, a spatiotemporal‑aware generative model tailored for massive online POI recommendation. By redesigning tokenization, incorporating hierarchical geographic indexing, and employing curriculum learning, the model improves hit‑rate and discovery‑rate metrics, laying a solid foundation for future downstream applications in Amap’s AI ecosystem.