- 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
- Which experiment supports "up to 87% Acc@1, exceeding training on real people"? If none, please remove it.
- Which number is correct for the count-supervision ablation, and which trait labeling is correct, Table 9 or Table 10?
- Can you test on traits that are never given as input or used during generation, and compare against the real person's values?
- 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?
- Does the public release include people derived from the Shanghai data? Do you have permission to release it?
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:
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.