acceptodds
Under review as a conference paper at ICLR 2027

BEYOND TRAJECTORY IMITATION: BEHAVIORALLY ALIGNED SYNTHETIC HUMAN MOBILITY GENERATION

Abstract

Existing synthetic mobility generators primarily optimize day-level trajectory realism, but realistic trajectories can still misrepresent how an individual moves. We study this missing objective as behaviorally aligned synthetic mobility generation: generating trajectories that realize an explicit multi-scale behavioral specification. We introduce CTGTrace, which represents individuals through 19 persistent characteristics and 16 daily intentions and realizes them through structured spatial generation, multi-day planning, and closed-loop trajectory decoding. Across five heterogeneous mobility populations and ten published generators, CTGTrace achieves a mean normalized request-realization error of 0.175 with correlation r = 0.925. We further show that the framework supports partial specifications, scenario transfer, and generation without an underlying real individual, and release the resulting synthetic dataset and code.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Human Mobility Generation, Mobility trajectory, Mobility behavioral features

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_6w6QOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper argues that fake mobility data should not just look realistic on average. Each fake person should also behave like a specific requested type of person. The user gives 35 behavior traits (19 long-term traits, such as how often someone goes back to the same places, and 16 daily goals), and the model, CTGTrace, generates a 10-day, hour-by-hour trajectory that matches them. It combines several parts: a home-location model, a layout generator for frequently visited places, a multi-day planner, a step-by-step decoder, a "progress tracker" that counts what is still left to do and pushes the decoder toward it, and a city-grounding step that places visits on real map cells. To create new fake people, it samples traits from a statistical model (a Gaussian copula) fitted on real people. It is tested on five datasets (one is private Shanghai phone data) against ten published generators. The paper reports small errors in matching requested traits, better population realism, and next-location prediction from fake data almost as good as from real data (0.542 vs 0.548). It also releases 1.16 million fake people.

Critical strengths

S1. A useful problem framing and diagnostic. Figure 1 clearly shows that published generators miss person-level behavior (the closest baseline is far from real-vs-real agreement). Turning this into an evaluation, with scores calibrated against real-vs-real agreement, is a good idea others could reuse.

S2. A broad evaluation. The paper covers five populations, ten baselines, and many types of checks: matching requested traits, population realism, usefulness for prediction, diversity and a privacy attack.

Critical weaknesses

W1. A headline claim has no supporting result. The contributions list says a model trained only on fake people predicts real people's next location "with up to 87% Acc@1, exceeding training on real people." No table shows this. Table 2 reports 0.542 for fake data vs 0.548 for real data, so fake is slightly worse, and the highest value anywhere in Table 17 is 0.616.

W2. The trait categories don't add up. Table 9 and Table 10

W3. The main results are largely built in.

  • The 35 traits used to score the model are the same ones given as input. The model also tracks the remaining targets and directly boosts the matching choices at every hour, so low error is expected rather than impressive.
  • Population realism is scored on the same traits, and new people's traits are sampled from a model fitted on those real traits.
  • The city-grounding step uses a visit-frequency map learned from the training data, which helps the cell-level scores.
  • Baselines get none of this help.

W4. Many baselines are broken. The appendix admits several adapted baselines "produce degenerate behavior." Table 2 shows coverage of 0.01–0.11 and cell scores of about 0.99 (near worst) for UPGLD, ControlTraj, Marionette and DiffTraj. Beating broken baselines proves little. On measures not built into the model, CTGTrace is still easy to tell apart from real data (classifier AUC 0.587 vs 0.487 for real-vs-real).

W5. Privacy. The only privacy evidence is one membership-inference attack, and the paper admits that request-based outputs keep more personal signal. There is no formal guarantee, yet it releases 1.16 million fake people, some derived from private Shanghai data.

Questions for authors
  1. Which experiment supports "up to 87% Acc@1, exceeding training on real people"? If none, please remove it.
  2. Which number is correct for the count-supervision ablation, and which trait labeling is correct, Table 9 or Table 10?
  3. Can you test on traits that are never given as input or used during generation, and compare against the real person's values?
  4. Can you give the working baselines the same fitted trait distribution and visit-frequency map, or remove them from CTGTrace, to see how much the generator itself contributes?
  5. Does the public release include people derived from the Shanghai data? Do you have permission to release it?
Reviewer_ibTdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies behaviorally aligned synthetic mobility generation, emphasizing that realistic aggregate trajectory statistics do not necessarily imply that each generated individual follows the intended behavioral specification. The proposed framework, CTGTrace, combines 19 persistent characteristics and 16 daily intentions with recurrent-place generation, multi-day planning, progress-aware trajectory decoding, and city-aware grounding. It supports complete and partial specifications, as well as de-novo generation through sampling behavioral characteristics from the training population. The evaluation covers five mobility populations and examines individual request realization, population fidelity, sequence realism, downstream utility, and component contributions through retrained ablations.

Critical strengths

1-The distinction between generating plausible trajectories and realizing a particular individual’s intended behavior is important. CTGTrace makes this objective concrete through persistent characteristics and daily intentions, and evaluates request–realization alignment alongside population-level fidelity. The separation of analytically imposed, explicitly tracked, and emergent quantities is particularly useful for understanding what the generator achieves. The same-request EPR comparison and the conditional Transformer controls further strengthen the argument that the structured generation process contributes beyond simply supplying behavioral information.

2-The evaluation across five populations extends beyond matching the requested characteristics. It includes additional sequence-level diagnostics, a learned two-sample test, spatial fidelity, and downstream models evaluated on held-out real data. Importantly, the main population comparison uses de-novo CTGTrace samples, avoiding dependence on behavioral requests from corresponding real test individuals. On next-visit prediction, the reported accuracy of 0.542 approaches the size-matched real-data reference of 0.548; occupancy prediction is also competitive.

The retrained ablations provide useful evidence about the mechanism. Removing recurrent-place count supervision produces substantial degradation, and removing trajectory-level progress consistently worsens alignment. The separate analysis of deterministic decoding corrections helps establish that the reported gains are not explained solely by late corrections. Together, these experiments make the empirical case reasonably convincing.

Critical weaknesses

1-The common ten-day, hourly city-cell representation enables consistent evaluation, but it also changes important aspects of several baselines’ native generation tasks. Some adapted implementations exhibit severe fidelity degradation or spatial collapse, making the resulting performance gaps difficult to attribute specifically to methodological differences.

Appendix C appropriately identifies MIRAGE, DITRAS, TimeGeo, and EPR as the primary comparisons and acknowledges the limitations of the more heavily transferred implementations. However, the main presentation still places substantial emphasis on comparison with all ten published generators. Representative native-versus-adapted checks, together with concrete training-budget or convergence evidence, would help establish how much behavior each implementation preserves through adaptation. The gains over the primary references remain persuasive, but the distinction between these comparisons and the transferred references should be clearer in the main text and tables.

2-CTGTrace combines learned components with substantial mobility-specific structure, including home/work/routine organization, analytical conversion of characteristics into targets, residual-demand corrections, and city-specific grounding. These are reasonable design choices, and the ablations provide evidence that learned recurrent-place usage and progress-aware decoding make meaningful contributions. Nevertheless, it remains unclear how much of the complete pipeline can be reused when the city, spatial support, or mobility patterns change.

The YJMob-B-to-YJMob-C experiment is encouraging, but it evaluates scenario transfer within the same metropolitan region and dataset family using target-scenario behavioral requests. As the paper correctly states, this is a narrower setting than cross-city transfer. The five-population evaluation establishes effectiveness across separately evaluated populations, while the data requirements and refitting effort for transferring the same generator to a new city remain unresolved.

Questions for authors

1-Can the authors provide representative checks showing how a baseline’s behavior changes when its native output is converted to the common representation, together with the relevant training-budget or convergence checks? It would also help to distinguish the primary comparisons from the more heavily adapted references directly in the main results.

2-In the YJMob-B-to-YJMob-C experiment, which components and statistics remain fixed, including the Home Location Model, feature normalization, spatial support, and destination-attractiveness prior? Are any quantities recomputed from the target population beyond supplying its behavioral requests? More generally, what target-city data and retraining or recalibration would be needed to deploy the same core generator in a new city?

Reviewer_1j4NOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a behavior-aligned synthetic movement trajectory generation task: the generated individual trajectories must not only appear realistic but also embody a clear, multiscale description of behavior. The authors claim that existing synthetic movement trajectory generators primarily focus on trajectory-level realism, and the generated individuals still exhibit significant discrepancies from real people in terms of behavioral characteristics such as exploration tendencies, regularity, and range of activity. To address this, the authors propose generating individual trajectories based on a multiscale behavioral description comprising persistent user characteristics and daily travel intentions, and designed the CTGTrace framework to achieve this goal. The framework first constructs a home-centered structure of frequently visited locations, then combines multi-day planning with hourly closed-loop decoding to generate trajectories. During decoding, it continuously tracks the gap between realized behavior and targets, and finally maps the generated locations onto a real-world urban grid. The paper compares the method across five mobility datasets, evaluating it in terms of both individual behavioral fidelity and data quality at the population level.

Critical strengths
  1. This paper highlights the shortcomings of existing generators from the perspective of individual behavior; the research motivation is clear and supported by data.
  2. Behavioral descriptions are categorized into two scales: persistent characteristics and daily intentions. Each dimension is clearly defined and supported by relevant research.
  3. The evaluation criteria are comprehensive, covering multiple aspects and perspectives, and most metrics use discrepancies in real-world data as a reference.
Critical weaknesses
  1. In the contributions, the authors states that 1.16 million synthetic individuals were released, but neither the main text nor the appendix explains how this dataset was generated or evaluated.
  2. The author states that models trained on synthetic data achieved an Acc@1 of up to 87% and outperformed those trained on real data, but the main text reports figures of 0.542 versus 0.548, and the highest value in the appendix is only 0.616. This claim lacks supporting evidence and contradicts the main text.
  3. The numerical values for the same result are inconsistent between the main text and the appendix. For example, the error increments for the Removing visitation-count supervision are listed as 0.588–0.686 and 0.560–0.637, respectively.
  4. The primary metric at the population level is the KS distance of 35 behavioral features, and since the model itself is generated based on these features, the evaluation inherently favors CTGTrace.
  5. The paper cites large-model-based methods such as LLMob, CoPB, and ELLMob, but does not list them as baselines.
Questions for authors

See Weaknesses

1 comment

Response to Reviewer 1j4N — Part 2/2

W4 — Feature KS overlaps conditioning.

Sec. 2 (p. 3) now outlines what the generator receives and produces. The concern does not hold for the de-novo population comparison (new Table 3, p. 9): the copula supplies only 19 persistent characteristics, never the 16 daily characteristics; the latter nevertheless reach 1.90×. Given identical copula samples, request-calibrated EPR scores 8.23× across 35 characteristics versus CTGTrace's 2.05× (new Table 23, p. 31). The sampled characteristics alone therefore cannot explain the gain. Independent tests not using the 35 characteristics include sequence realism (1.42× versus MIRAGE 3.58×) and a learned two-sample classifier (new Tables 24–25, p. 31).

We fixed seven held-out traits and their definitions in code before evaluation (new Table 16, p. 28). None is among the 35 or computable from the 19 inputs: 24-hour return rate, day-to-day similarity, weekday–weekend contrast, exploration growth exponent, visitation Zipf exponent, trip-timing burstiness, daily motif profile. Mean KS relative to real-vs-real, across five populations (new Table 21, p. 30): real 1.11; CTGTrace 2.11; MIRAGE 4.91; DITRAS/TimeGeo/EPR 8.0–9.2. For individuals, CTGTrace matches all seven requester traits more closely than the real person nearest in the 19 input traits (new Table 2, p. 7). See Sec. 5.1 (p. 7) and App. E.5 (pp. 28–30).

W5 — LLM baselines.

We added LLMob (NeurIPS 2024) with authors’ code/prompts and format-only changes. The retired GPT-3.5-turbo-0613 is replaced by the open-weight gpt-oss-20b, run locally at the same temperature (App. C.2 and new Table 9, p. 21). CTGTrace beats LLMob on every new Table 3 metric:

ModelFeature KS ↓Realism ↓Next Acc@1 ↑Occ. R² ↑Person gap ↓
CTGTrace2.051.42.542.909.09
LLMob9.7810.8.394.731.49

CoPB requires demographics, intention labels and intention-typed places; ELLMob requires textual societal-event descriptions. None exists in our five populations. Without these they reduce to history-conditioned agents, which is represented by LLMob. LLM agents are also costly and hard to control: LLMob needs about 250 model calls per synthetic person, against one decoding run for CTGTrace, and nothing checks its generated days against requested characteristics (App. C.2, p. 21).

The revision fixes missing documentation, inconsistencies, and clarity; new held-out-trait and LLMob experiments strengthen our claims. Roadmaps now open Sec. 5 (p. 6) and the appendix (p. 13). We kindly ask the reviewer to reconsider the score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.