acceptodds
Under review as a conference paper at ICLR 2027

Plan-to-Synthesis: Cross-City Human Mobility Generation via Semantic Latent Flow Matching

Abstract

Human mobility generation aims to synthesize realistic point-of-interest (POI) visitation trajectories and has become an important tool for travel behavior modeling, transportation management, and urban planning. Existing diffusion-based methods achieve high fidelity but require per-city generation, given the inherent heterogeneity of geospatial locations and POI categories, while large language model-based methods generalize across cities but remain too costly at scale, especially for long-horizon trajectory generation. To address this, we propose SeMoFlow, a Semantic human Mobility generation framework based on latent Flow matching. We first encode heterogeneous POIs from different cities into a shared cross-city representation space via hierarchical Semantic IDs, where shared prefixes capture transferable semantics, and successive codes progressively refine the representation toward individual POIs. Building on the semantic IDs, SeMoFlow follows a plan-to-synthesis hierarchical generation paradigm, in which an autoregressive planner generates coarse-grained semantic and recurrence patterns, and a flow matching realizer synthesizes fine-grained suffix latents. The generated latents are subsequently decoded and grounded to concrete POIs. Extensive experiments on large-scale multi-city datasets show that SeMoFlow achieves higher trajectory fidelity than existing baselines, preserves city-specific mobility motifs, and supports both joint multi-city generation and effective cross-city transfer.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Urban Mobility, Flow matching, Trajectory Generation, Transportation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6iX7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SeMoFlow, a framework for generating realistic human mobility trajectories across different cities. It uses hierarchical Semantic IDs to represent POIs in a shared semantic space, an autoregressive planner to capture activity patterns and repeated visits, and a flow-matching model to generate fine-grained locations. Experiments on two multi-city datasets show improvements in trajectory fidelity and demonstrate the potential for cross-city transfer.

Critical strengths
  1. The paper addresses an important challenge in human mobility generation: transferring mobility patterns across cities with different POI vocabularies and spatial structures.

  2. The hierarchical Semantic ID representation provides a reasonable way to capture shared activity semantics while preserving fine-grained POI information.

  3. The plan-to-synthesis framework explicitly considers recurring visits and separates coarse-grained activity planning from fine-grained location generation.

Critical weaknesses
  1. The methodological novelty is somewhat limited, as the framework mainly combines existing techniques, including Semantic IDs, autoregressive planning, and flow matching.

  2. Although computational efficiency is part of the motivation, the paper does not provide runtime, memory usage, or generation throughput comparisons to support its scalability claims.

  3. Cross-city transfer is evaluated on only one unseen city, Dallas, making it difficult to assess how well the approach generalizes to more diverse urban environments.

  4. The ablation studies are limited to Chicago and do not compare flow matching with alternative suffix generation methods, leaving its specific contribution insufficiently validated.

  5. The evaluation mainly relies on distribution-level metrics, without demonstrating whether the generated trajectories improve downstream applications or preserve detailed individual-level behavior.

  6. The final POI grounding stage uses a fixed first-candidate rule when multiple POIs share the same Semantic ID, but the potential impact of these collisions on generation diversity is not investigated.

Questions for authors
  1. How are the POI transition-graph embeddings constructed for Dallas in the zero-shot transfer setting, given that no target-city mobility observations are available?

  2. Can the authors compare the flow-matching realizer with simpler suffix generation alternatives under matched computational budgets and report actual generation efficiency to better justify this design choice?

13 comments
AuthorsOpenReview ↗

Response to Reviewer 6iX7 (Part 1/4)

Thank you for the valuable comments and suggestions. Our responses are as follows:

Weakness 1: The methodological novelty clarification.

Response: Thank you for the valuable comments. Our contribution is that we design a new semantic human mobility generation framework that effectively integrates hierarchical semantic tokenization with a plan-to-synthesis coarse-to-fine design for cross-city generation. To address the key challenge that POI vocabularies are incompatible across cities, which makes multi-city joint training hard to support, we encode POIs from multiple cities into a unified representation space through hierarchical Semantic IDs, so that POIs from different cities can be modeled within the shared SID space. Then, leveraging the hierarchy of semantic tokenization, we design a plan-to-synthesis coarse-to-fine generative model to seperate the transferable activity layer from the city-specific POI layer at the architecture level, which enables cross-city transferability and has the potential to serve as a basis for larger-scale POI mobility foundation models.

Weakness 2: Computational efficiency comparisons to support its scalability claims..

Response: Thank you for the comment. We add a sampling-efficiency comparison between the flow-matching realizer and an autoregressive (AR) realizer of the same size (9.2M parameters) trained under the same compute budget. We sample 3,000 weekly trajectories with batch size 512 on an idle H200 GPU, and report the median sampling time over five repetition runs and the throughput (events/s):

Suffix generatorSampling time (s) ↓Throughput (events/s) ↑
AR10.576,609
Flow matching, 20 steps3.0123,234
Flow matching, 5 steps1.0864,831

The results show that the flow matching with 5 steps is faster than the AR realizer. We have added this comparison to the appendix C.3.

AuthorsOpenReview ↗

Response to Reviewer 6iX7 (Part 2/4)

Weakness 3: Cross-city transfer is evaluated on only one unseen city, Dallas.

Response: Thank you for the valuable comment. To further evaluate cross-city generalization, we added two additional target cities, i.e., Miami and Seattle, using the same four source cities for pretraining as in the Dallas experiment.
In both cities, cross-city fine-tuning outperforms training from scratch on most metrics when target-city data are limited. With 5–10% of the data, fine-tuning is better on travel distance, radius of gyration, and recurrence in both cities, while training from scratch is better on POI frequency in Miami at 10%. As more target data become available, the gap narrows. These trends are consistent with the Dallas results.

Miami (JSD ↓; FT = cross-city fine-tuning, Sc = per-city training from scratch; 0% exists only for FT)

Metric0% FT1% FT1% Sc5% FT5% Sc10% FT10% Sc20% FT20% Sc30% FT30% Sc40% FT40% Sc50% FT50% Sc100% FT100% Sc
JSD_POI0.51200.24580.23620.15950.16150.14950.12510.13480.11290.12900.10710.12940.10170.13160.10290.12420.0929
JSD_geo0.17310.02090.01350.00800.01180.00580.00880.00440.00770.00330.00650.00310.00500.00330.00400.00310.0029
JSD_travel0.14880.12550.30770.02980.11120.03940.05660.03510.03070.03600.02780.03410.02060.03410.02320.03340.0210
JSD_gyr0.09950.03420.26710.02260.14440.02410.05460.01960.01460.01990.01100.01880.00760.01710.01020.02010.0081
JSD_motif0.13230.02980.03750.00970.01040.00800.00850.00660.01000.00680.00670.00650.00490.00690.00660.00650.0071
JSD_recur0.26510.03050.02230.00170.00970.00120.00550.00150.00400.00150.00270.00210.00210.00180.00220.00190.0026

Seattle

Metric0% FT1% FT1% Sc5% FT5% Sc10% FT10% Sc20% FT20% Sc30% FT30% Sc40% FT40% Sc50% FT50% Sc100% FT100% Sc
JSD_POI0.50060.26410.24220.13980.15650.11690.12720.10320.11080.09620.10060.09280.08970.09740.08950.08730.0788
JSD_geo0.16760.02440.02150.01050.01690.00670.01010.00600.00860.00470.00590.00350.00530.00420.00540.00370.0045
JSD_travel0.12250.13510.39390.04300.12490.01610.06330.01600.03370.01510.03690.01290.02180.01660.02000.01320.0136
JSD_gyr0.09750.02190.23130.00470.07230.00510.02690.00440.00840.00560.02000.00360.00660.00430.00420.00300.0046
JSD_motif0.09480.09880.06080.02410.02390.00400.00830.00340.01010.00270.00710.00290.00540.00330.00510.00270.0057
JSD_recur0.20000.16660.07520.00620.04810.00170.01910.00180.01580.00200.00680.00170.00630.00170.00440.00190.0049

We have included these additional results for these two cities in Appendix C.1 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer 6iX7 (Part 3/4)

Weakness 4 and Question 2: The ablation studies are limited to Chicago and do not compare flow matching with alternative suffix generation methods

Response: Thank you for the suggestions. We extend the component ablations in Table 2 from Chicago to Boston, New York City, and Pittsburgh, with the same variants: w/o Graph removes the POI transition graph embeddings from SID tokenization; w/o Router removes the Explore-Retrieve mechanism from the planner; and w/o Planner generates the full SID through latent flow matching without prefix planning. Values are means over three sampling seeds with 30,000 generated weeks each. The results are as follows:

Boston

VariantJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
w/o Graph0.08280.00660.18110.23830.00070.0021
w/o Router0.07760.00900.01380.01020.00320.0051
w/o Planner0.35120.04610.11390.31590.00030.0042
SeMoFlow0.07960.00780.01210.00460.00100.0023

New York City

VariantJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
w/o Graph0.14490.00790.22230.27110.00230.0030
w/o Router0.14020.00780.04240.01440.00700.0079
w/o planner0.45730.04850.28670.53470.00080.0069
SeMoFlow0.14430.00850.03790.01730.00270.0036

Pittsburgh

VariantJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
w/o Graph0.06360.00700.17660.24310.00210.0010
w/o Router0.06400.00850.04100.03640.00450.0068
w/o planner0.31370.04960.29120.62090.00090.0034
SeMoFlow0.06480.00800.03960.03940.00310.0017

Consistent with Chicago, removing the graph embeddings sharply increases travel-distance and radius JSD in all three cities, and removing the router increases recurrence JSD in all three cities. Removing the planner decreases the performance in all three cities.

Alternative suffix generation methods: We conduct experiments to compare the alternatives of the suffix generation module in this paper with an autoregressive generator as suggested. Both have the same size (9.2M parameters) and the same training budget. Regarding generation quality on Chicago, the results are as follows:

Suffix generatorJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
AR0.09840.00550.00720.00090.00060.0038
SeMoFlow (flow matching, 20 steps)0.09430.00520.01720.00180.00250.0020
Flow matching, 5 steps0.09530.00530.01740.00190.00290.0022

The flow-matching realizer is better than the AR realizer on POI, geohash, and recurrence, while AR is better on travel distance, radius of gyration, and motif. The two are therefore comparable in quality, and flow matching is chosen for its much higher sampling efficiency, as shown in the response to Weakness 2.

Reviewer_BNnLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SeMoFlow addresses city-conditioned generation of anonymous POI visitation trajectories. It combines semantic, transition-graph, and geographic features into shared hierarchical Semantic IDs through residual quantization. A recurrence-aware autoregressive planner generates time gaps and SID prefixes; conditional flow matching jointly generates suffix representations, which are decoded against the target-city POI catalog. Experiments cover two data sources, multiple cities, component ablations, and transfer to Dallas at different target-data budgets.

Critical strengths
  1. The method addresses incompatible POI vocabularies across cities through a coherent combination of shared representations, hierarchical planning, and catalog-constrained generation. The ablations support the usefulness of these components.
  2. The evaluation provides meaningful positive evidence: device-disjoint splits, strong recurrence and motif results on Veraset, improvements from joint training over the per-city variant on Foursquare, and a matched-budget Dallas comparison showing benefits from source-city learning when target data are limited.
Critical weaknesses
  1. The proposed semantic hierarchy and its incremental transfer value are not sufficiently distinguished from generic coarse-to-fine compression and multi-city pretraining. Residual quantization does not itself assign activity semantics to the first three codes. Figure 5 provides useful Chicago evidence, but POI-weighted majority-category purity necessarily cannot decrease as prefix groups are refined, so group size matters. The choice of the three-code boundary and cross-city semantic consistency remain insufficiently justified. Dallas compares transfer with the same model trained from scratch, rather than an alternative shared or transfer model. Joint training also has higher POI JSD than the per-city variant in all four Veraset cities.

  2. The reported gains need stronger interpretation in terms of baseline comparability, decoding contributions, and behavioral validity. Baseline adaptation to a common concrete-POI output space is insufficiently specified, particularly given the large MobiDiff gaps. SeMoFlow's POI advantage over MoveSim is substantially smaller and reverses in NYC. Equation 8 also uses a legitimate training-frequency prior whose contribution is not isolated. Finally, all six main metrics compare distributions: several contain trajectory structure, but matching them separately does not establish joint temporal/spatial consistency within synthetic trajectories. Recurrence measures visit-index gaps rather than elapsed time. These are limitations of the evidence, not proof of incorrect implementations or invalid generation.

  3. I think MoveSim is paper in published about 5 or more years ago, and the MobiDiff is published in this year, but I noted that the latest MobiDiff even worse than MoveSim in some indicators, did the author consider the adaptive for the task when doing the evaluation and baseline implementations to confirm comparable?

Questions for authors
  1. Clarification: What motivates the three-code prefix boundary, and how are transition-graph features and the new city embedding obtained in zero-shot Dallas, where the appendix explicitly excludes target mobility observations and visitation priors?

  2. Clarification: How were MoveSim and MobiDiff adapted to weekly trajectories and concrete POI identities, including location vocabularies, decoding/postprocessing, and relevant sampling and tuning settings?

  3. Additional evidence, if available: Do group-size-aware or cross-city analyses support the prefix's activity interpretation? Do matched-budget shared-model comparisons or controlled suffix-generator comparisons distinguish the proposed design from generic pretraining, and how much of the POI advantage remains without the frequency prior? Such evidence would materially affect attribution of the gains; I am not requesting every possible ablation.

6 comments
AuthorsOpenReview ↗

Response to Reviewer BNnL (Part 1/3)

Thank you for the valuable comments and suggestions. We group related sub-points in the weaknesses and questions together and respond point by point below.

Weakness 1, Question 1, and Question 3: The choice of the three-code boundary. POI-weighted majority-category purity necessarily cannot decrease as prefix groups are refined. Group-size aware to support prefix's activity interpretation.

Response: Thanks for your comments and suggestions. We agree with the reviewer's comment that residual quantization does not itself assign activity semantics to the first three codes. The three-code boundary is a design choice based on our observation that the first three codes are sufficient to determine a dominant POI category.

Fig. 5(c) gives an example in which a three-code prefix corresponds to a dominant POI category and the remaining codes separate individual POIs. We agree that the purity in Fig. 5(b) is affected by group size: as the prefix becomes longer, each group contains fewer POIs, and the share of the dominant category goes up, so purity increases.

Following the reviewer's suggestion, we add a group-size-aware metric to support the activity interpretation of the prefix. Following existing SID interpretation work (LMIndexer, ICML 2024 [1]; GSID, EMNLP 2025 [2]), we report the adjusted mutual information (AMI, Vinh et al., 2009 [3]) between prefix groups and POI categories. AMI measures how much the prefix tells us about the category, after subtracting what a random assignment with the same group sizes would already achieve. Unlike purity, it is not affected by group size. We compute AMI only over prefix groups containing at least two labeled POIs, since a single-POI group carries no information about category agreement. The results are as follows:

CityDepth 1Depth 2Depth 3Depth 4Depth 5Depth 6
Boston0.810.540.370.240.220.25
Chicago0.820.540.350.230.220.24
NYC0.750.570.410.250.200.21
Pittsburgh0.810.610.390.280.280.33
Dallas0.770.560.430.260.220.24

This result shows that the category information is carried mainly by the first several codes. For the deeper codes, the groups become very small, and there is little category information left for the suffix codes to add, and they mainly serve to separate individual POIs.

[1] Jin et al., Language Models as Semantic Indexers. ICML 2024.

[2] Yang et al., GSID: Generative Semantic Indexing for E-Commerce Product Understanding. EMNLP 2025.

[3] Vinh et al. Information theoretic measures for clusterings comparison: is a correction for chance necessary? ICML. 2009.

We also conduct a prefix-depth ablation on Chicago, where the planner generates the first P codes and the realizer generates the remaining codes:

Prefix depth PJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
10.2820 ± 0.00030.0220 ± 0.00010.3600 ± 0.01060.4922 ± 0.01700.0266 ± 0.00030.1454 ± 0.0052
20.1744 ± 0.00030.0087 ± 0.00010.0588 ± 0.00080.0444 ± 0.00030.0136 ± 0.00020.0044 ± 0.0003
3 (ours)0.0944 ± 0.00010.0053 ± 0.00010.0170 ± 0.00050.0025 ± 0.00070.0025 ± 0.00010.0021 ± 0.0001
40.1047 ± 0.00010.0062 ± 0.00000.0148 ± 0.00050.0110 ± 0.00020.0041 ± 0.00010.0107 ± 0.0001
50.1044 ± 0.00030.0065 ± 0.00020.0102 ± 0.00030.0081 ± 0.00050.0036 ± 0.00010.0145 ± 0.0007
6 (planner only)0.1054 ± 0.00030.0071 ± 0.00010.0133 ± 0.00000.0110 ± 0.00060.0040 ± 0.00010.0178 ± 0.0004

A one-code or two-code prefix is worse on almost all metrics, since such short prefixes are not enough to determine the activity type. A three-code prefix gives the best results on POI, geohash, radius of gyration, motif, and recurrence. Extending the prefix to four or more codes does not improve most divergences and even degrades most metrics, because the prefix groups become much smaller and revisits have to be modeled over a far sparser vocabulary. The three-code prefix therefore gives the best balance between activity-level planning and fine-grained realization.

AuthorsOpenReview ↗

Response to Reviewer BNnL (Part 2/3)

Question 1: Clarification of how are transition-graph features and the new city embedding obtained in zero-shot Dallas

Response: Thank you for the question. In the zero-shot setting, no Dallas mobility data are available, so we build a heterogeneous graph from catalog information only: Dallas POIs are connected to their POI categories and to the census block groups (CBGs) they are located in, with reverse edges included. We initialize the graph encoder with the weights trained on the source cities, set the target-specific node ID embeddings to zero, and obtain the Dallas graph embeddings by a single forward pass. The frozen tokenizer then combines these graph embeddings with the semantic features and the city-relative geographic features (i.e., the offset of each POI from the city centre) to assign SIDs to Dallas POIs. No visitation data from Dallas are used in this setting. For the few-shot fine-tuning setting, the transition graph is built from the budgeted target-city training trajectories and the graph encoder is fine-tuned on it.

Weakness 1, Question 3: Distinguish transfer value of semantic hierarchy, coarse-to-fine compression and multi-city pretraining. Do matched-budget shared-model comparisons or controlled suffix-generator comparisons distinguish the proposed design from generic pretraining.

Response: Thank you for the comments and suggestions. We conduct two comparisons: (i) SID-AR: keep the same hierarchical SIDs and replace the plan-to-synthesis with a planner that autoregressively generates the full six-code SID, and (ii) compare with MoveGPT [1], which represents locations by grid cells with continuous embeddings instead of Semantic IDs. The results are as follows:

ModelJSD_POIJSD_geoJSD_travelJSD_gyrJSD_motifJSD_recur
SID-AR (same SIDs)0.12780.01910.13040.17670.00090.0050
MoveGPT (no SIDs)0.22360.14050.00560.02310.00640.0063
SeMoFlow (cross-city fine-tuning)0.12060.00460.01970.01290.00540.0013

SeMoFlow outperforms SID-AR on 5 of the 6 metrics; SID-AR is better only on JSD_motif. Compared with MoveGPT, SeMoFlow is better on all metrics except JSD_travel. A likely reason is that MoveGPT feeds the normalized coordinates of each location directly into its location embedding and predicts the next location by similarity search over all candidate locations, so the spatial displacement of every step is conditioned on continuous coordinates. In SeMoFlow, coordinates only enter the tokenizer and the final POI is chosen within the SID bucket during grounding. This adds position noise at each visit.

[1] Han, Chonghua, et al. "MoveGPT: Scaling Mobility Foundation Models with Spatially-Aware Mixture of Experts." arXiv preprint arXiv:2505.18670.

Weakness 1, Question 3: Joint training also has higher POI JSD than the per-city variant in all four Veraset cities.

Response: Thank you for the comments. Thank you for pointing out this observation. We agree that joint training has higher POI JSD than the per-city variant. We believe this is because a per-city model decodes over a much smaller POI vocabulary than the joint model trained on four cities, which makes POI-level decoding easier. The goal of our design is to support multi-city joint training and cross-city transfer. The experiments show that the joint model achieves better or comparable results to the per-city models on most metrics, which indicates that the Semantic ID design enables building a unified pretrained model across multiple cities. In addition, Fig. 4 shows that transferring from this pretrained model to a new city outperforms training from scratch when target-city data are limited.

AuthorsOpenReview ↗

Response to Reviewer BNnL (Part 3/3)

Weakness 2, Question 3: Training-frequency prior's contribution is not isolated.

Response: Thank you for the comments and suggestions. We add an ablation on Chicago that removes the training-frequency prior from the SID decoding process. The results are as follows:

VariantJSD_POI ↓JSD_geo ↓JSD_travel ↓JSD_gyr ↓JSD_motif ↓JSD_recur ↓
w/o frequency prior0.09990.00520.01740.00310.00260.0021
SeMoFlow0.09430.00520.01720.00180.00250.0020

The frequency prior contributes mainly to POI-level fidelity.

Weakness 2, Weakness 3, Question 2: Baseline (Movesim and Mobidiff) adaptation and performance comparison clarification.

Response: Thank you for the comments and suggestions. The observation that MobiDiff trails MoveSim on some metrics is not specific to our setting. In MobiDiff’s own paper, on their check-in data (arXiv 2607.08357, Tab. 2), MoveSim beats MobiDiff on travel distance and radius of gyration in all three cities, and the authors describe MoveSim as “the strongest utility baseline”. In our setting, MoveSim performs well on the two spatial marginal metrics, POI frequency and geohash distribution (JSD 0.09–0.14 and 0.01–0.02 across the four cities), because its next-location likelihood directly fits how often each POI is visited. On the metrics that depend on trajectory structure, however, it is far behind MobiDiff and the other generative baselines on travel distance, radius of gyration, and recurrence metrics (Table 1).

We include MobiDiff because it is a masked discrete diffusion model over trajectory tokens, which differs from the latent-diffusion-based baselines and is therefore worth comparing against. Since its code is not public, we implemented it strictly following the paper. However, the large discrete location vocabulary in our setting (69k–394k POIs per city, versus 2.8k–5.2k in its paper) may cause the per-step categorical prediction of the discrete diffusion model to spread probability mass thinly and concentrate samples on a small subset of POIs, leading to weaker POI and geohash distribution results.

We give the detailed adaptation of two methods as follows:

  • MoveSim. We replaced its original cell-tower/grid-based location vocabulary with one token for each POI observed in the training split. To accommodate weekly trajectories, each device’s weekly POI sequence was resampled to 48 positions for training, retaining the original model’s input length. At our POI vocabulary sizes, the original model’s dense location-relation matrices were prohibitively expensive. We therefore used a simplified variant based on its released recurrent generator, trained with next-location likelihood. During generation, we directly sampled sequences with the average weekly trajectory length in the training data. At each subsequent sampling step, we sampled a POI token from the model’s predicted next-location distribution and converted it to the corresponding POI ID.
  • MobiDiff. We implemented its multi-channel discrete diffusion framework based on the paper. To generate concrete POI identities, we used the catalog POI ID as the fine-grained location channel, with geohash-5 cells as the coarse-grained location channel. Each training sequence corresponded to one device-week. We replaced time-of-day with hour-of-week (168 bins) and represented inter-visit gaps using 64 bins. Trajectory lengths were sampled from the training length distribution. During decoding, the fine-grained location token directly identified a POI. Its coarse region and category were then obtained from the POI catalog to ensure consistency across channels. We use the paper’s 20-step confidence-based unmasking with temperature 1 and top-k 20, and the paper’s architecture and training schedule, with the best-validation checkpoint.

We will clarify these adaptations and settings in Appendix B.5.

Weakness 2: Joint temporal/spatial consistency within synthetic trajectories and elapsed time-aware Recurrence. Response: Thank you for the comments and suggestions. We added two metrics, the JSD over the joint distribution of visit hour and location and a time-aware recurrence metric based on the elapsed hours between returns to the same POI. On the Veraset cities, SeMoFlow achieves the best time–location JSD in all four cities. On time-aware recurrence, SeMoFlow achieves the best result in NYC and the second-best result in the other three cities. We have added the two metrics to Tables 1 and 8 in the revised paper.

Reviewer_fDyVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose SeMoFlow, a Semantic human Mobility generation framework based on latent Flow matching. They generate semantic IDs of point-of-interest (POI) and use an autoregressive planner to generate trajectories over prefixes of the semantic IDs and then flow-matching to generate the suffix of the semantic IDs.

Critical strengths
  1. The paper is well written and easy to read and follow.
  2. The idea of the hierarchical semantic IDs is nice and appropriate.
  3. Using a planner to do the high level trajectory generation and flow-matching for the details is also nice.
  4. The evaluation shows that the method performs better than the baselines. They also show that cross-city learning improves the result.
Critical weaknesses
  1. It seems that the concept of a POI is very generic, what are the limitations?
  2. Line 428: Is the target-city POI catalogue produced by jointly learning tokenizers across a set of cities in which the target city is included? If yes, then this is a (reasonable) limitation of the approach and should be stated clearly. If not, how compatible is the POI catalogue with the learned joint one?
  3. It is hard to judge how the method generalises from five cities.
  4. Overall the paper is nice, but it is unclear to me whether the method and the results are significant enough.
Questions for authors
  1. How much data do you need from a city to get sufficient quality? Maybe look at the scaling behaviour of real data.
  2. Fig 4, how do you explain that the quality goes down with more real data?
  3. Would it make sense to have some information about the cities, such as a map? Now I understand that you only take some POI trajectories as input.
  4. What about making the trajectories conditional on type of person? It seems that the POIs will vary depending on whether a person works, if (s)he has kids, if they live down town or in the suburbs, etc.
5 comments
AuthorsOpenReview ↗

Response to Reviewer fDyV (Part 1/3)

Thank you for the valuable comments and suggestions. Our responses are as follows:

Weakness 1 and Question 3: The concept of a POI is generic. Would it make sense to have some information about the cities, such as a map?

Response: Thank you for the valuable comments. We use POI visit sequences to represent human mobility because they directly reflect residents’ visitation activity behaviour, and the POI categories provide information about the types of activities associated with these visits. The POI visiting trajectories are widely used in activity-level mobility behavior analysis. We agree with the reviewer's comments that the concept of POI has some limitations; for example, it does not describe the transportation modes or routes between visits, or lack the city map, or layout contexts.

We appreciate the suggestion to include map information, which is a great idea. Map information, road networks, and urban layout could provide additional conditions for our generation framework. This information can also be treated as physical constraints to help filter out candidate POIs that cannot be reached within the generated time gap, given an assumed travel mode. We have clarified the current limitations and discuss these extensions in the revised paper.

Weakness 2: Is the target-city POI catalogue produced by jointly learning tokenizers?

Response: Thank you for the valuable comments. The target-city catalog is not encoded by a tokenizer that includes the target city. The tokenizer (RQ-VAE weights and all six 48-entry codebooks) is fitted only on Boston, Chicago, NYC, and Pittsburgh and then frozen. When performing cross-city transfer, the POI catalog of the target city, e.g., Dallas, is encoded by a single forward pass through the frozen tokenizer, and the pretrained codebooks are unchanged.

Compatibility with the learned joint space: The target catalog is compatible with the joint tokenizer because the tokenizer operates on POI attributes rather than POI identities. Dallas POIs are described by the same category taxonomy, the same city-relative geographic representation, and the same graph encoder as the source cities, so their feature vectors lie in the same feature space on which the codebooks were learned. Some existing work in generative recommendation has also shown that assigning SIDs to unseen items through a frozen tokenizer is effective (TIGER[1]; Singh et al.[2]), and that shared discrete codes can transfer across domains (VQ-Rec[3]).

In our setting, each POI enters the tokenizer as a feature vector with three parts: (i) semantic features from POI metadata and text (category, name), which carry no city-specific mobility information; (ii) geographic features expressed relative to the city centre, which for a new city are derived from the catalog coordinates alone; and (iii) the graph embedding of the POI transition/distance graph. In the zero-shot setting, no Dallas mobility data are available, so we build a heterogeneous graph from catalog information only: Dallas POIs are connected to their POI categories and to the census block groups (CBGs) they are located in, with reverse edges included. No visitation data from Dallas are used in this setting. At 1–100% we build the transition graph from the budgeted target-city training trajectories and fine-tune the graph encoder on it.

[1] Rajput, Shashank, et al. "Recommender systems with generative retrieval." Advances in Neural Information Processing Systems (2023).

[2] Singh, Anima, et al. "Better generalization with semantic IDs: A case study in ranking for recommendations." Proceedings of the 18th ACM Conference on Recommender Systems. 2024.

[3] Hou, Yupeng, et al. "Learning vector-quantized item representation for transferable sequential recommenders." Proceedings of the ACM Web Conference 2023.

AuthorsOpenReview ↗

Response to Reviewer fDyV (Part 2/3)

Weakness 3: How the method generalises from five cities.

Response: Thank you for the valuable comment. To further evaluate cross-city generalization, we added two additional target cities, i.e., Miami and Seattle, using the same four source cities for pretraining as in the Dallas experiment.
In both cities, cross-city fine-tuning outperforms training from scratch on most metrics when target-city data are limited. With 5–10% of the data, fine-tuning is better on travel distance, radius of gyration, and recurrence in both cities, while training from scratch is better on POI frequency in Miami at 10%. As more target data become available, the gap narrows. These trends are consistent with the Dallas results.

Miami (JSD ↓; FT = cross-city fine-tuning, Sc = per-city training from scratch; 0% exists only for FT)

Metric0% FT1% FT1% Sc5% FT5% Sc10% FT10% Sc20% FT20% Sc30% FT30% Sc40% FT40% Sc50% FT50% Sc100% FT100% Sc
JSD_POI0.51200.24580.23620.15950.16150.14950.12510.13480.11290.12900.10710.12940.10170.13160.10290.12420.0929
JSD_geo0.17310.02090.01350.00800.01180.00580.00880.00440.00770.00330.00650.00310.00500.00330.00400.00310.0029
JSD_travel0.14880.12550.30770.02980.11120.03940.05660.03510.03070.03600.02780.03410.02060.03410.02320.03340.0210
JSD_gyr0.09950.03420.26710.02260.14440.02410.05460.01960.01460.01990.01100.01880.00760.01710.01020.02010.0081
JSD_motif0.13230.02980.03750.00970.01040.00800.00850.00660.01000.00680.00670.00650.00490.00690.00660.00650.0071
JSD_recur0.26510.03050.02230.00170.00970.00120.00550.00150.00400.00150.00270.00210.00210.00180.00220.00190.0026

Seattle

Metric0% FT1% FT1% Sc5% FT5% Sc10% FT10% Sc20% FT20% Sc30% FT30% Sc40% FT40% Sc50% FT50% Sc100% FT100% Sc
JSD_POI0.50060.26410.24220.13980.15650.11690.12720.10320.11080.09620.10060.09280.08970.09740.08950.08730.0788
JSD_geo0.16760.02440.02150.01050.01690.00670.01010.00600.00860.00470.00590.00350.00530.00420.00540.00370.0045
JSD_travel0.12250.13510.39390.04300.12490.01610.06330.01600.03370.01510.03690.01290.02180.01660.02000.01320.0136
JSD_gyr0.09750.02190.23130.00470.07230.00510.02690.00440.00840.00560.02000.00360.00660.00430.00420.00300.0046
JSD_motif0.09480.09880.06080.02410.02390.00400.00830.00340.01010.00270.00710.00290.00540.00330.00510.00270.0057
JSD_recur0.20000.16660.07520.00620.04810.00170.01910.00180.01580.00200.00680.00170.00630.00170.00440.00190.0049

We have included these additional results in these two cities in Appendix C.1 of the revised paper.

AuthorsOpenReview ↗

Response to Reviewer fDyV (Part 3/3)

Question 1: How much data do you need from a city to get sufficient quality?

Response: Thank you for the suggestion. To evaluate how performance scales with the amount of target-city data, we evaluate more data fractions, particularly in the low-data regime: 0.1%, 0.5%, 1%, 2%, 5%, 10%, 20%, 30%, 40%, 50%, and 100%. We conducted these experiments on three held-out cities, Dallas, Miami, and Seattle, comparing fine-tuning the jointly trained source model with training a per-city model from scratch under the same amount of target-city data.

To quantify sufficient quality, we use the per-city model trained on 100% of the target-city training data as the reference, and report the smallest evaluated data fraction at which each JSD of the fine-tuned model is no more than 10% above its reference value. Under this criterion, cross-city transfer reaches the full-data level on travel distance and recurrence with 5–10% of the target-city data, and on motif with 10–20%, in all three cities. Geohash density needs more data (40% in Seattle, 100% in Dallas, and slightly above the threshold in Miami at 100%).

Question 2: The quality goes down with more real data?

Response: Thank you for pointing this out. After carefully checking the results, We found that with little target-city data (e.g., less than 5%), models in both settings tend to generate repeated daily patterns and trajectories with less diverse lengths.

The daily motif metric is particularly sensitive to this, because each day is summarized as a small directed graph of the visited locations, which depends jointly on how many places are visited per day and the order of revisits. In addition, the real daily motif distribution is concentrated on a small number of motifs, so small changes in the generated patterns cause noticeable fluctuations in JSD_motif, although its absolute range across budgets is small (about 0.008). With multiple sampling runs, the large fluctuation with increasing real data no longer appears. In most low-data sampling rates, cross-city fine-tuning outperforms training from scratch. We have updated Fig. 4 accordingly.

Question 4: What about making the trajectories conditional on type of person?

Response: Thank you for the valuable comment and suggestions. We agree that mobility patterns vary with personal attributes such as age, employment, and other personal features. We did not model this because the Veraset data are anonymized for privacy and do not contain individual persona or profile information. Our framework can support such conditioning: persona tokens can be added as a condition to both the autoregressive planner and the flow-matching realizer. Some travel survey datasets do provide demographic labels, but they are usually small in scale. Fine-tuning our multi-city pretrained model on such small labelled datasets would enable persona-conditioned generation and analysis. We have discussed this in the new discussion section (section 6) of the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.