- 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
-
The paper addresses an important challenge in human mobility generation: transferring mobility patterns across cities with different POI vocabularies and spatial structures.
-
The hierarchical Semantic ID representation provides a reasonable way to capture shared activity semantics while preserving fine-grained POI information.
-
The plan-to-synthesis framework explicitly considers recurring visits and separates coarse-grained activity planning from fine-grained location generation.
Critical weaknesses
-
The methodological novelty is somewhat limited, as the framework mainly combines existing techniques, including Semantic IDs, autoregressive planning, and flow matching.
-
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.
-
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.
-
The ablation studies are limited to Chicago and do not compare flow matching with alternative suffix generation methods, leaving its specific contribution insufficiently validated.
-
The evaluation mainly relies on distribution-level metrics, without demonstrating whether the generated trajectories improve downstream applications or preserve detailed individual-level behavior.
-
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
-
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?
-
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
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)
| Metric | 0% FT | 1% FT | 1% Sc | 5% FT | 5% Sc | 10% FT | 10% Sc | 20% FT | 20% Sc | 30% FT | 30% Sc | 40% FT | 40% Sc | 50% FT | 50% Sc | 100% FT | 100% Sc |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JSD_POI | 0.5120 | 0.2458 | 0.2362 | 0.1595 | 0.1615 | 0.1495 | 0.1251 | 0.1348 | 0.1129 | 0.1290 | 0.1071 | 0.1294 | 0.1017 | 0.1316 | 0.1029 | 0.1242 | 0.0929 |
| JSD_geo | 0.1731 | 0.0209 | 0.0135 | 0.0080 | 0.0118 | 0.0058 | 0.0088 | 0.0044 | 0.0077 | 0.0033 | 0.0065 | 0.0031 | 0.0050 | 0.0033 | 0.0040 | 0.0031 | 0.0029 |
| JSD_travel | 0.1488 | 0.1255 | 0.3077 | 0.0298 | 0.1112 | 0.0394 | 0.0566 | 0.0351 | 0.0307 | 0.0360 | 0.0278 | 0.0341 | 0.0206 | 0.0341 | 0.0232 | 0.0334 | 0.0210 |
| JSD_gyr | 0.0995 | 0.0342 | 0.2671 | 0.0226 | 0.1444 | 0.0241 | 0.0546 | 0.0196 | 0.0146 | 0.0199 | 0.0110 | 0.0188 | 0.0076 | 0.0171 | 0.0102 | 0.0201 | 0.0081 |
| JSD_motif | 0.1323 | 0.0298 | 0.0375 | 0.0097 | 0.0104 | 0.0080 | 0.0085 | 0.0066 | 0.0100 | 0.0068 | 0.0067 | 0.0065 | 0.0049 | 0.0069 | 0.0066 | 0.0065 | 0.0071 |
| JSD_recur | 0.2651 | 0.0305 | 0.0223 | 0.0017 | 0.0097 | 0.0012 | 0.0055 | 0.0015 | 0.0040 | 0.0015 | 0.0027 | 0.0021 | 0.0021 | 0.0018 | 0.0022 | 0.0019 | 0.0026 |
Seattle
| Metric | 0% FT | 1% FT | 1% Sc | 5% FT | 5% Sc | 10% FT | 10% Sc | 20% FT | 20% Sc | 30% FT | 30% Sc | 40% FT | 40% Sc | 50% FT | 50% Sc | 100% FT | 100% Sc |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| JSD_POI | 0.5006 | 0.2641 | 0.2422 | 0.1398 | 0.1565 | 0.1169 | 0.1272 | 0.1032 | 0.1108 | 0.0962 | 0.1006 | 0.0928 | 0.0897 | 0.0974 | 0.0895 | 0.0873 | 0.0788 |
| JSD_geo | 0.1676 | 0.0244 | 0.0215 | 0.0105 | 0.0169 | 0.0067 | 0.0101 | 0.0060 | 0.0086 | 0.0047 | 0.0059 | 0.0035 | 0.0053 | 0.0042 | 0.0054 | 0.0037 | 0.0045 |
| JSD_travel | 0.1225 | 0.1351 | 0.3939 | 0.0430 | 0.1249 | 0.0161 | 0.0633 | 0.0160 | 0.0337 | 0.0151 | 0.0369 | 0.0129 | 0.0218 | 0.0166 | 0.0200 | 0.0132 | 0.0136 |
| JSD_gyr | 0.0975 | 0.0219 | 0.2313 | 0.0047 | 0.0723 | 0.0051 | 0.0269 | 0.0044 | 0.0084 | 0.0056 | 0.0200 | 0.0036 | 0.0066 | 0.0043 | 0.0042 | 0.0030 | 0.0046 |
| JSD_motif | 0.0948 | 0.0988 | 0.0608 | 0.0241 | 0.0239 | 0.0040 | 0.0083 | 0.0034 | 0.0101 | 0.0027 | 0.0071 | 0.0029 | 0.0054 | 0.0033 | 0.0051 | 0.0027 | 0.0057 |
| JSD_recur | 0.2000 | 0.1666 | 0.0752 | 0.0062 | 0.0481 | 0.0017 | 0.0191 | 0.0018 | 0.0158 | 0.0020 | 0.0068 | 0.0017 | 0.0063 | 0.0017 | 0.0044 | 0.0019 | 0.0049 |
We have included these additional results for these two cities in Appendix C.1 of the revised paper.
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
| Variant | JSD_POI ↓ | JSD_geo ↓ | JSD_travel ↓ | JSD_gyr ↓ | JSD_motif ↓ | JSD_recur ↓ |
|---|---|---|---|---|---|---|
| w/o Graph | 0.0828 | 0.0066 | 0.1811 | 0.2383 | 0.0007 | 0.0021 |
| w/o Router | 0.0776 | 0.0090 | 0.0138 | 0.0102 | 0.0032 | 0.0051 |
| w/o Planner | 0.3512 | 0.0461 | 0.1139 | 0.3159 | 0.0003 | 0.0042 |
| SeMoFlow | 0.0796 | 0.0078 | 0.0121 | 0.0046 | 0.0010 | 0.0023 |
New York City
| Variant | JSD_POI ↓ | JSD_geo ↓ | JSD_travel ↓ | JSD_gyr ↓ | JSD_motif ↓ | JSD_recur ↓ |
|---|---|---|---|---|---|---|
| w/o Graph | 0.1449 | 0.0079 | 0.2223 | 0.2711 | 0.0023 | 0.0030 |
| w/o Router | 0.1402 | 0.0078 | 0.0424 | 0.0144 | 0.0070 | 0.0079 |
| w/o planner | 0.4573 | 0.0485 | 0.2867 | 0.5347 | 0.0008 | 0.0069 |
| SeMoFlow | 0.1443 | 0.0085 | 0.0379 | 0.0173 | 0.0027 | 0.0036 |
Pittsburgh
| Variant | JSD_POI ↓ | JSD_geo ↓ | JSD_travel ↓ | JSD_gyr ↓ | JSD_motif ↓ | JSD_recur ↓ |
|---|---|---|---|---|---|---|
| w/o Graph | 0.0636 | 0.0070 | 0.1766 | 0.2431 | 0.0021 | 0.0010 |
| w/o Router | 0.0640 | 0.0085 | 0.0410 | 0.0364 | 0.0045 | 0.0068 |
| w/o planner | 0.3137 | 0.0496 | 0.2912 | 0.6209 | 0.0009 | 0.0034 |
| SeMoFlow | 0.0648 | 0.0080 | 0.0396 | 0.0394 | 0.0031 | 0.0017 |
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 generator | JSD_POI ↓ | JSD_geo ↓ | JSD_travel ↓ | JSD_gyr ↓ | JSD_motif ↓ | JSD_recur ↓ |
|---|---|---|---|---|---|---|
| AR | 0.0984 | 0.0055 | 0.0072 | 0.0009 | 0.0006 | 0.0038 |
| SeMoFlow (flow matching, 20 steps) | 0.0943 | 0.0052 | 0.0172 | 0.0018 | 0.0025 | 0.0020 |
| Flow matching, 5 steps | 0.0953 | 0.0053 | 0.0174 | 0.0019 | 0.0029 | 0.0022 |
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.
Response to Reviewer 6iX7 (Part 1/4)
Thank you for the valuable comments and suggestions. Our responses are as follows:
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.
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):
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.