Spacetime-GR: Generative POI Model

• Spacetime-GR is a generative model that leverages spatiotemporal context and hierarchical POI indexing to address online point-of-interest recommendation challenges.
• The model employs a modified Transformer with multimodal POI embeddings and post-training adaptation to support both ranking and end-to-end generation tasks.
• Empirical results show significant gains in AUC, CTR, and CVR on large industrial datasets, demonstrating scalability to hundreds of millions of users and POIs.

Searching arXiv for the target paper and closely related recommendation context. {"query":"(Lin et al., 22 Aug 2025) Spacetime-GR A Spacetime-Aware Generative Model for Large Scale Online POI Recommendation", "max_results": 5} Searching for recent related papers on generative recommendation and POI recommendation for contextual verification. {"query":"generative recommendation POI recommendation arXiv 2025 2026", "max_results": 10} Spacetime-GR is a spacetime-aware generative model for large-scale online Point-of-Interest (POI) recommendation in which “GR” denotes Generative Recommendation rather than General Relativity. It was introduced to address a specific open problem: although Generative Recommendation has gradually assumed a dominant position in recommendation tasks such as video and product recommendation, POI recommendation remains difficult because user preferences are significantly affected by spatiotemporal variations. The model extends the sequence modeling ability of generative architectures by combining a geographic-aware hierarchical POI indexing strategy, a spatiotemporal encoding module, multimodal POI embeddings, and post-training adaptation strategies that support multiple output formats and downstream application scenarios. The reported system was evaluated on public benchmark datasets and large-scale industrial datasets, and was described as the first generative model deployed in online POI recommendation services that scale to hundreds of millions of POIs and users (Lin et al., 22 Aug 2025).

1. Problem formulation and design objective

Spacetime-GR is designed for online POI recommendation under conditions in which the recommendation target is not determined by interaction history alone. In the formulation presented for the model, user action sequences are strongly conditioned by temporal context and geographic context, and those two signals must therefore be represented as first-class modeling objects rather than appended side information. This is the meaning of “spacetime-aware” in the system: the model incorporates flexible spatiotemporal information encoding directly into the generative sequence model (Lin et al., 22 Aug 2025).

The architecture is also motivated by scale. The paper treats large vocabulary modeling as a central systems problem because industrial POI platforms must represent extremely large candidate spaces. A flat-token formulation over hundreds of millions of POIs would make generation and scoring inefficient. Spacetime-GR therefore addresses POI recommendation as a joint sequence-modeling and vocabulary-structuring problem, with the goal of supporting both offline evaluation and practical online deployment.

A further design objective is deployment flexibility. Rather than restricting the model to one inference interface, the post-training scheme is built to produce embeddings, ranking scores, and POI candidates. This permits use in both ranking and end-to-end recommendation, and it supports recall, reranking, and direct generation within one framework.

2. Input structure, hierarchical indexing, and spatiotemporal encoding

A central architectural component is the geographic-aware hierarchical POI indexing strategy. Each POI $p_i$ is represented by two tokens: $block_i$, the spatial “block” containing $p_i$, and $inner_i$, the unique index of $p_i$ within its block. The mapping is $(block_i, inner_i) \leftrightarrow p_i$. In the reported formulation, this reduces the softmax search space from $O(100\text{M})$ to $O(400\text{K})$ while also improving spatial generalization (Lin et al., 22 Aug 2025).

Each user action $s_i$ is encoded as a tuple $(u_i, block_i, inner_i, a_i)$. Here $block_i$0 is the user’s spatiotemporal context token, including discretized time $block_i$1 and user geo-location $block_i$2; $block_i$3 and $block_i$4 are the hierarchical POI tokens; and $block_i$5 is the action type token. Temporal features are discretized into month, weekday, day, and hour, while spatial features are hashed. Each feature type is assigned its own embedding layer.

The input embeddings are defined as follows: $block_i$6

$block_i$7

$block_i$8

$block_i$9

In this construction, $p_i$0 are normalized dynamic weights, and the embedding layers correspond to time, geo-location, POI or block identity, category, and action type. A defining feature of the model is that spatiotemporal context is directly encoded as sequence tokens. The paper contrasts this with previous work that treated space and time as side information, and argues that the tokenized formulation allows self-attention to capture interactions among time, space, and actions more explicitly.

3. Backbone architecture and pre-training regime

The backbone of Spacetime-GR is a modified Llama 2 12-layer decoder-only Transformer. At each step, prediction is produced through a language-modeling head followed by softmax: $p_i$1 This places the model within the decoder-only generative paradigm while preserving the specialized POI tokenization and spatiotemporal input structure (Lin et al., 22 Aug 2025).

Pre-training is performed on action sequences using a cross-entropy objective over hierarchical POI tokens only for interest-based actions: $p_i$2 where $p_i$3 is a binary interest-action indicator. This objective ties the generative task directly to the hierarchical POI representation rather than to a flat POI vocabulary.

The training procedure also uses curriculum learning. Training begins on simpler single-pattern sequences, specifically local, pre-travel, and in-travel patterns, and later proceeds to multi-pattern data intended to reflect more realistic user behaviors. This sequencing is presented as a way to stabilize learning under heterogeneous mobility and intent patterns.

4. Multimodal enrichment and post-training adaptation

Spacetime-GR augments sequence modeling with multimodal POI embeddings. In post-training, POI representations are enriched with textual attributes such as names, addresses, tags, and reviews, together with images describing visual characteristics. A pre-trained multimodal LLM produces pooled multimodal POI embeddings, which are injected into the $p_i$4 token embedding for downstream tasks (Lin et al., 22 Aug 2025).

The post-training stage includes three supervised fine-tuning or alignment strategies. The first is embedding-based ranking SFT, which adopts a dual-tower architecture. User context and POI are encoded separately as $p_i$5 and $p_i$6, and the model is trained with a contrastive objective: $p_i$7

The second is generative ranking SFT. In this variant, user context and candidate POIs are jointly encoded in the Transformer, with an attention mask ensuring per-POI independence for candidates. A sigmoid-activated binary classification score is predicted for each candidate through the $p_i$8 token state: $p_i$9

$inner_i$0

The third is an alignment stage based on Direct Preference Optimization (DPO): $inner_i$1 where $inner_i$2 is the finetuned model, $inner_i$3 the frozen reference model, $inner_i$4 the sigmoid, and $inner_i$5 a scaling factor. The paper states that DPO can be guided by spatiotemporal-aligned preference data.

These adaptation stages are intended to support three output modalities: embeddings, ranking scores, and top-$inner_i$6 POI candidates. The supported application scenarios are ranking and end-to-end recommendation.

5. Empirical evaluation and deployment characteristics

The industrial-scale evaluation uses data from Amap. Pre-training uses 578M samples with average sequence length approximately 146, totaling 77B actions. SFT and alignment use 31M training samples with average sequence length approximately 301 and approximately 11 POI candidates per request. The reported infrastructure is 96 H20 GPUs for 7 days in pre-training and 16 H20 GPUs for 2 days in finetuning (Lin et al., 22 Aug 2025).

For ranking evaluation, the reported metrics are AUC, CTR, and CVR; for POI recall and ranking, Hit Rate@$inner_i$7 is used. On the industrial ranking task, AUC improves from 0.7043 for the online baseline to 0.7229 for embedding SFT and 0.7272 for generative SFT, and the combined system achieves 0.7385. In live traffic split, the model reports $inner_i$8 CTR and $inner_i$9 CVR, described as statistically significant. For generative recommendation with DPO alignment, the paper reports outperformance over the online system in LLM and human assessments, including a 67% “win” rate at system level and 69.9% at POI level. Reported online latency is 10–20 ms per recommendation.

On public datasets, the model is evaluated on Foursquare-NYC, Foursquare-TKY, and Gowalla-CA. The reported comparison with STHGCN is:

Dataset	STHGCN	Spacetime-GR
Foursquare-NYC	0.2734	0.2920
Foursquare-TKY	0.2950	0.2610
Gowalla-CA	0.1730	0.1659

The accompanying interpretation in the paper is that Spacetime-GR sets a new state-of-the-art on NYC and is competitive on the others, despite not using other user sequences as in STHGCN.

6. Ablations, practical scope, and interpretive significance

The ablation study attributes measurable performance degradation to removing major architectural components. Removing spatiotemporal encoding yields $p_i$0 percentage points hr@1 and $p_i$1 percentage points hr@100. Removing hierarchical POI indexing also degrades performance. Omitting curriculum learning yields slightly lower performance, while adding multimodal embeddings contributes $p_i$2 percentage points to $p_i$3 percentage points AUC. The paper also reports that longer user sequence length improves results up to sequence length approximately 128, and that the model has a higher discovery rate, recommending more novel POIs than the baseline (Lin et al., 22 Aug 2025).

In practical terms, the system is presented as a unified generative recommendation framework for POI services at national scale. The reported deployment to real-time industrial systems with hundreds of millions of users and POIs, together with output support for embeddings, scores, and generated candidate lists, indicates that the model is intended not merely as an offline research architecture but as a production recommender substrate.

A common source of ambiguity is the model name itself. In this context, “spacetime” refers to spatiotemporal variations in user behavior and POI context, not to physical spacetime. The technical contribution lies in representing these variations inside the sequence model rather than outside it. This suggests a more general recommender principle: for domains in which user intent is strongly conditioned by where and when actions occur, direct token-level integration of spatiotemporal context can be more natural than auxiliary-feature injection. A plausible implication is that Spacetime-GR serves as a template for extending generative recommendation from comparatively stationary catalog settings to strongly context-conditioned mobility and local-service environments.