GeoGen: A Two-stage Coarse-to-Fine Framework for Fine-grained Synthetic Location-based Social Network Trajectory Generation

AI summary

Overview

• Research area: Generative AI for spatio-temporal data — specifically synthetic location-based social network (LBSN) check-in trajectory generation, sitting at the intersection of diffusion models, sequence-to-sequence learning, and urban mobility mining.
• Technical level: Advanced. The paper assumes familiarity with denoising diffusion probabilistic models, U-Net architectures, Transformer encoder-decoder models, and neural temporal point processes.
• Scope: The paper proposes GeoGen, a two-stage coarse-to-fine framework that generates fine-grained (minute-level) synthetic LBSN check-in trajectories, evaluated on four real-world datasets against eight baselines on fidelity and utility.

What This Paper Is About

LBSN check-in trajectories are valuable for recommendation, advertising, urban planning, and pandemic response, but collecting them at scale is expensive and raises privacy concerns. The goal is to generate synthetic check-in trajectories that mimic real data's spatio-temporal statistics while protecting privacy — a hard problem because check-in data is spatially discrete (POIs form a finite set of locations, unlike GPS coordinates) and temporally irregular (uneven time gaps and variable trajectory lengths), whereas existing diffusion-based trajectory generators were designed for continuous, fixed-interval, fixed-length GPS data.

Key Contributions

1. Problem framing: The paper explicitly targets the gap that prior trajectory diffusion models (DiffTraj, ControlTraj) assume spatially continuous, temporally regular, fixed-length trajectories, which is incompatible with LBSN check-in data.
2. Stage 1 — S²TDiff: A Sparsity-aware Spatio-temporal Diffusion model whose denoising network, SASG-UNet (Spatially-Aware Sparsely-Gated U-Net), uses hierarchical 1D convolution blocks instead of standard 2D convolutions and an S²G Attention module in the skip connections to handle sparse intensity and diverse movement rates.
3. Stage 2 — Coarse2FineNet: A Transformer-based Seq2Seq architecture with a POI Context-aware Encoder (containing a Dynamic Context Fusion mechanism over a shared POI embedding codebook) and a Multi-task Hybrid-Head Decoder that predicts the next POI and its fine-grained timestamp, using a neural temporal point process for timing instead of direct regression.
4. Empirical validation: Experiments on four real-world datasets (FS-NYC, FS-TKY, FS-ATX, GW-STO) against eight baselines, plus ablations, a granularity/efficiency study, and a downstream next check-in prediction task.

Main Findings

• Fidelity on FS-TKY: GeoGen improves over the second-best method by over 69% in the Distance metric and 55% in the Radius metric on the FS-TKY dataset. Its FS-TKY scores are Distance 0.015, Radius 0.043, Interval 0.081, Length 0.120, Average 0.065 (lower JSD is better).
• Best average across all four datasets: GeoGen achieves the lowest average JSD in every dataset — FS-NYC 0.085, FS-TKY 0.065, FS-ATX 0.096, GW-STO 0.176 — versus the next-best averages of 0.140 (DiffTraj on FS-NYC), 0.140 (ControlTraj on FS-TKY), 0.190 (ControlTraj on FS-ATX), and 0.226 (MoveSim on GW-STO).
• Length on GW-STO: The paper reports a 38% improvement in the Length metric on GW-STO (GeoGen 0.080 vs. MoveSim 0.138 and ControlTraj 0.141).
• Baseline weaknesses: Rule-based methods (SMM, TimeGEO, Hawkes) rely on fixed statistical assumptions and generally underperform. DiffTraj is competitive on spatial metrics for dense urban data like FS-NYC (Radius 0.085) but poor on Length (0.320) due to no awareness of discrete POIs; on FS-TKY its Length JSD is 0.744. ControlTraj models POIs as road segments, which is incompatible with temporally irregular data.
• Utility (next check-in prediction): Training a predictor on synthetic data and testing on real data, GeoGen gives the closest-to-real performance: FS-NYC RMSE 0.225 / ED 9.21 and FS-TKY RMSE 0.352 / ED 15.82, versus real-data references of 0.187 / 7.01 and 0.235 / 13.70, and versus SeqGAN (0.426 / 16.34 and 0.689 / 23.12), MoveSim (0.371 / 10.28 and 0.671 / 16.90), DiffTraj (0.424 / 15.27 and 0.728 / 18.28) and ControlTraj (0.321 / 14.29 and 0.491 / 16.28).
• Ablation on FS-NYC: Replacing interpolation with replication (S²TDiff-Rep: 0.084/0.071/0.074/0.128) or a fixed central point (S²TDiff-Fix: 0.097/0.073/0.079/0.130) hurts; removing S²G Attention (0.087/0.069/0.080/0.135) degrades both spatial and temporal metrics; removing Dynamic Context Fusion (C2F w/o DCF: 0.119/0.138/0.096/0.169) causes the largest drop. Complete GeoGen scores 0.079/0.065/0.067/0.127.
• Granularity trade-off (GW-STO): Smaller intervals (2–4 hours) give better temporal fidelity (Interval), while spatial metrics (Distance, Radius) are best at a 6-hour interval (Distance 0.073, Radius 0.154). Moving from a 2-hour to a 6-hour interval cuts Stage 1 memory from 2.36 GB to 0.35 GB and raises Stage 1 throughput from 11,322 to 50,540 samples/sec (more than fourfold). Stage 2 memory stays comparatively stable (17.55 GB at 2 hours down to 3.62 GB at 8 hours) with throughput rising modestly from 39 to 49 samples/sec.
• Distributional fit: CDF curves on FS-TKY (for distance, radius, length, and interval) show GeoGen's generated distributions most closely align with the original data.
• Efficiency: The two-stage design explicitly resolves the trade-off between computational efficiency and data quality; Stage 1 uses batch size 256 and Stage 2 uses batch size 16 in the granularity experiments.

Methodology in Plain English

GeoGen splits the hard problem into two easier ones.

Stage 1 — make the data look like something a diffusion model can handle. Raw check-in trajectories are sparse and unevenly spaced. GeoGen chops the time axis into fixed-length slots of length I (e.g., 6 hours), giving L = [D/I] slots. For each slot it computes a mean geographic coordinate and an "intensity" equal to the number of check-ins in that slot. Empty slots are filled by linear interpolation between the nearest preceding and following observed slots, and boundary gaps are filled by treating the trajectory as a loop (circular interpolation). This turns a ragged, discrete event stream into a smooth, regular, continuous signal.

A diffusion model (S²TDiff) then learns to generate these latent movement sequences. Its denoising network, SASG-UNet, is a U-Net whose convolutions are replaced by hierarchical 1D blocks with multi-scale average pooling, so it can capture patterns at several resolutions cheaply. In the skip connections sits the S²G Attention module, with two pathways: an Intensity-Gated Pathway that learns a sigmoid gate to emphasize slots that actually contain check-ins (instead of interpolated ones), and a Spatially-Aware Attention Pathway that injects a bias derived from the Haversine distance between consecutive noisy points, so the model has a direct signal about plausible movement speeds.

Stage 2 — turn the smooth coarse signal back into discrete check-ins with realistic timing. Coarse2FineNet is a Transformer encoder-decoder. The encoder first filters the latent sequence by an intensity threshold γ, keeping only slots that had real activity, then embeds each point spatially (a linear layer) and temporally (Time2Vec, which captures both linear and cyclical time patterns). A shared POI embedding codebook encodes each candidate POI's coordinates, temporal visit-frequency vector, and category. A Dynamic Context Fusion module runs two attention computations — a spatial one weighting POIs by geographic proximity (with temperature τ_s) and a temporal one weighting POIs by visit-frequency similarity to a projected time query — then combines them with learned softmax fusion weights β_s and β_t. The fused POI context is added back to the point features via a residual connection and passed through the Transformer encoder.

The decoder works autoregressively. Given the encoder output and the check-ins generated so far, a hidden state feeds two heads: a linear layer producing POI logits, and a neural temporal point process with intensity λ(t) = Softplus(W_time h + b_time) for the timestamp. The timestamp is sampled by inverse transform sampling, t = F⁻¹(u) with u ~ Uniform(0,1), which is what lets the model produce realistic non-uniform time gaps rather than assuming a fixed interval.

Training. The model is trained with teacher forcing and causal masking, minimizing a weighted sum of three losses: cross-entropy for POI prediction, a temporal point process negative log-likelihood for timing, and a spatial consistency loss penalizing the Euclidean distance between predicted and ground-truth POI coordinates. The weights are learnable.

Why This Matters

Impact on research. GeoGen is a concrete demonstration that diffusion models — previously bound to continuous, fixed-interval GPS data — can be adapted to the discrete and irregular structure of LBSN check-ins through a coarse-to-fine decomposition. It shows that the "compress to a regular latent grid, then decode back to discrete events" pattern is a workable general recipe, and it establishes a new state of the art on four widely used datasets with a released codebase.

Real-world applications:

• POI recommendation: Synthetic trajectories that preserve mobility patterns let recommenders be trained and stress-tested without exposing real user traces.
• Advertising and business location selection: Planners can simulate foot-traffic and visit patterns across candidate sites using realistic synthetic populations.
• Urban mobility analysis and transportation planning: Agencies can model demand and movement flows without purchasing or collecting sensitive location data.
• Pandemic intervention: Epidemiologists can simulate contact and mobility patterns for outbreak modeling, an application the paper explicitly cites.

Industry relevance. Companies that hold check-in data (social platforms, mapping services, retail analytics) face both legal restrictions and user trust costs around sharing it. A generation pipeline that produces faithful, useful synthetic substitutes offers a path to share data and collaborate externally while the paper frames this as safeguarding privacy. The reported efficiency gains — over fourfold higher Stage 1 throughput and roughly a 6.7x memory reduction when moving from a 2-hour to a 6-hour interval — matter directly for whether such generation can be run at the large scales the paper targets.

Future Directions

• Explicit privacy quantification. The paper motivates synthetic data as a privacy safeguard, but the provided content reports no privacy metrics (e.g., membership inference or attribute inference resistance) or privacy guarantees. Measuring how much privacy synthetic trajectories actually provide, and against what adversaries, is an open question.
• Per-dataset granularity selection. The granularity study shows spatial fidelity peaks at a 6-hour interval while temporal fidelity peaks at 2–4 hours. An adaptive or learned scheme for choosing I per dataset, rather than a single global setting, would better balance quality and cost.
• Cross-city and cross-platform generalization. Evaluation is limited to four datasets (three Foursquare, one Gowalla). Whether the framework transfers to other platforms, non-urban regions, or datasets with different sparsity levels is untested.
• Stage 2 efficiency and POI vocabulary scaling. Stage 2 memory is roughly 17.55 GB at a 2-hour interval and throughput stays in the tens of samples per second. Whether the dual-attention Dynamic Context Fusion scales to POI vocabularies and city sizes far beyond the four studied datasets is not reported.

Target Audience

Researchers and practitioners working on generative modeling for spatio-temporal data, synthetic data generation, urban computing, and human mobility analysis. It is most valuable to readers already comfortable with diffusion models and Transformer architectures who want to understand how to bridge continuous generative models and discrete, irregularly sampled event data. Privacy engineers and data-sharing officers at organizations holding LBSN data will also find the framing relevant, though they should note that the provided content does not report privacy metrics.

Authors’ abstract

Location-Based Social Network (LBSN) check-in trajectory data are important for many practical applications, like POI recommendation, advertising, and pandemic intervention. However, the high collection costs and ever-increasing privacy concerns prevent us from accessing large-scale LBSN trajectory data. The recent advances in synthetic data generation provide us with a new opportunity to achieve this, which utilizes generative AI to generate synthetic data that preserves the characteristics of real data while ensuring privacy protection. However, generating synthetic LBSN check-in trajectories remains challenging due to their spatially discrete, temporally irregular nature and the complex spatio-temporal patterns caused by sparse activities and uncertain human mobility. To address this challenge, we propose GeoGen, a two-stage coarse-to-fine framework for large-scale LBSN check-in trajectory generation. In the first stage, we reconstruct spatially continuous, temporally regular latent movement sequences from the original LBSN check-in trajectories and then design a Sparsity-aware Spatio-temporal Diffusion model (S$^2$TDiff) with an efficient denosing network to learn their underlying behavioral patterns. In the second stage, we design Coarse2FineNet, a Transformer-based Seq2Seq architecture equipped with a dynamic context fusion mechanism in the encoder and a multi-task hybrid-head decoder, which generates fine-grained LBSN trajectories based on coarse-grained latent movement sequences by modeling semantic relevance and behavioral uncertainty. Extensive experiments on four real-world datasets show that GeoGen excels state-of-the-art models for both fidelity and utility evaluation, e.g., it increases over 69% and 55% in distance and radius metrics on the FS-TKY dataset.