acceptodds
Under review as a conference paper at ICLR 2027

TRANSFERRING ACTIVITY PROTOTYPES FOR ZERO- SHOT CROSS-CITY MOBILITY GENERATION

Abstract

preserving privacy. It has become increasingly important for urban planning and management in smart cities. However, training a generative model requires abundant local real trajectory data, which many cities cannot provide. An effective alternative is to transfer behavioral knowledge from data-rich cities. Nevertheless, differences in location semantics and spatial structure hinder the direct reuse of source individuals’ activity patterns. To address this challenge, we propose SHAPE, a cross-city mobility generation framework that extracts shared activity patterns as core representations for cross-city transfer. The key idea of SHAPE is to encode individual activity patterns into interpretable, city-independent activity prototypes with residual finite scalar quantization, and transfer them to the target city with its static map and the spatial distribution of its residents, thus generating high-quality individual mobilities with adaptation to the target urban environment. In zero-shot cross-city mobility generation, we compare SHAPE with three families of state-of-the-art baselines on four source-to-target cases. SHAPE ranks first on four of the five metrics and at least second on the remaining one. In an oracle test from New York City to Paris, SHAPE uses no individual trajectory of Paris and reaches 90.1% of the OD CPC of an oracle given those trajectories. The extraction of shared activity patterns, rather than reliance on abundant target-city ground truth, is the primary driver of performance improvements. By introducing a city-independent codebook of activity prototypes, our framework effectively balances the trade-off between model capacity and cross-city transfer robustness.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Zero-Shot Transfer, Human Mobility Generation, Discrete Representation Learning, Urban Computing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_sp5NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies zero-shot cross-city mobility trajectories generation tasks, transferring the human activity sequences of individuals in one city to another. It proposes SHAPE, which first construct alignment between source and target city, then use FSQ to encode each activity into discrete activity prototype, finally use prototypes to guide location selection. Experiments on four cases show SHAPE ranks first on multiple metrics and reaches 90.1% of an oracle’s OD CPC.

Critical strengths

The motivation of this paper is well defined, focusing on a practical problem. It separates the activity schedule from its spatial placement and obtaining discrete activity prototypes, using intervening opportunities to make movement comparable across cities. This makes the transfer both well-motivated and interpretable. Besides, The experiments are comprehensive, covering four transfer directions, three families of baselines, and five metrics at both the aggregate and individual levels, SHAPE performs consistently well across diverse directions.

Critical weaknesses
  1. The paper claims the extraction of shared activity patterns is the key to performance improvements, but in table 2,removing Stage 3 reduces OD CPC to 0.3412, whereas removing the Stage 2 reduces it to 0.4875 (full model: 0.5806). Moreover, the importance of stages may vary across metrics? The paper should explain how these metrics relate to each other or which one best reflects the claimed contribution.

  2. SHAPE copies the activity type of each source activity and uses it to select cells with matching POIs, so it matches the POI-category distribution by construction. The largest reported margin (25%–90.9% on POI-category JSD) may reflect the use of this input rather than more realistic trajectories.

Questions for authors
  1. The paper states that the model accesses the target city only through its static map and residential distribution, but the reproducibility statement mentions target-domain “distance statistics” used during generation. What exactly are they?

  2. How were the Stage 3 hyperparameters (ring edges, temperatures, top-k, etc.) selected, and was any target-city data involved?

  3. Stage 1 alone “generates no mobility by itself”, yet Table 2 reports metrics for “Stage 1 only” and “Stage 2 only.” How are trajectories generated in the settings without Stage 3?

  4. COLA, which the paper cites as motivation, is a related cross-city trajectory model; although it may require some target-city data, a discussion of why it cannot be adapted to this setting (or a comparison under a few-shot variant) would help position the work. TimeGeo and DITRAS share the same schedule-then-location split that SHAPE adopts (Sec. 3.1), and would be mechanistic baselines.

  5. Sec. 5 and Figure 2 state that L1 is tied to interval/jump length and L2 to activity type (via L_sem), but Figure 7 and Table 6 show the opposite: L1 mainly refines activity type (0.835 bits vs. 0.131 bits on jump length), while L2 adds only 0.020 bits on activity type and instead refines time and movement. In addition, A.4 states that the hierarchy emerges “without layerwise semantic injection,” whereas L_sem supervises each layer with a different feature group (A.2.2). Which description is correct?

Reviewer_dmUDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper tries to generate realistic daily movement of people in a new city without any movement data from that city ("zero-shot"). It learns from check-in data in one city and uses only the new city's map and where people live. It works in three steps. First, it matches map areas across the two cities by what kinds of places are there. Second, it turns each person's activities into short codes ("activity prototypes"). Third, it places those activities in the new city using a "how many opportunities are nearby" rule plus a distance penalty. The person's daily schedule (times and activity types) is copied directly from the source city; only the locations are new. The paper tests on New York, Paris and California and reports winning 4 of 5 metrics, reaching 90.1% of an "oracle" on New York to Paris.

Critical strengths

S1. Clean design with a full ablation. Splitting "what and when people do things" from "where they do them in the new city" is a sensible idea. The paper tests all 7 combinations of its parts (Table 2), and those numbers are consistent with the text.

S2. Useful goal. Generating mobility for cities with no movement data matters for planning, and the paper compares against several types of baselines (rule-based, regression, diffusion).

Critical weaknesses

W1. It is not really zero-shot. Section 4.1 says the new city's home distribution is "estimated from the training days of the target city," meaning from that city's check-in data. The reproducibility statement also says target-city "distance statistics" are used during generation. But the oracle analysis (Section 4.3) treats Paris distance statistics as extra information SHAPE does not have. The paper contradicts itself about what target-city data it uses, so the 90.1%-of-oracle result can't be trusted.

W2. The biggest win is matched by design. Section 4.2 presents the POI-category score as SHAPE's largest improvement (median 60.7%). But Section 4.3 admits SHAPE "matches [it] by construction" because it re-weights using Paris's place counts, and it scores 265% of the oracle. Matching a metric by design is not evidence that behavior transfers.

W3. The distance part conflicts with the paper's own claim. The paper says distances are "not comparable between cities" and that placement does not depend on the source city's geography. Yet the placement score (Eq. 12) uses a distance penalty in kilometres, with its strength (γ) and minimum distance (d_min) fitted on the source city.

W4. The "interpretable codes" story doesn't match the numbers. Appendix A.2.2 says each code layer is trained to capture a different feature, but Appendix A.4 says the layers organize themselves "without layerwise semantic injection." Section 5 says layer 0 = time, layer 1 = distance, layer 2 = activity type. Table 6 shows something else: activity type is almost fully captured by layer 0, layer 2 adds almost nothing, and layer 1 mostly adds activity-type information, not distance. Figure 4 shows only 8 hand-picked codes out of 64.

W5. Copying the schedule inflates scores. Since times and activity types are copied directly from the source city, part of the time-related score comes for free. On jump distance, the one measure in kilometres, SHAPE reaches only 2.8% of the oracle (Section 4.3), which goes against the claim that behavior transfers.

Questions for authors
  1. Which target-city numbers come from target-city check-ins (home distribution, distance statistics, home locations)? Can you rerun the main table using only the map or census data? Does the 90.1% result still hold?
  2. Where do the target-city place counts come from? If they come from the same check-in dataset, that is another leak of target-city movement data.
  3. How does the kilometre distance term (fitted on the source city) fit with the "independent of source geography" claim? What happens if you remove it (γ = 0)?
  4. Which is correct, A.2.2 (each layer supervised separately) or A.4 (no layerwise supervision)? How do you explain Table 6?
  5. Why are COLA and GTG not compared?
  6. How many fake users are generated compared with the real test users, and how is the main metric adjusted for that?
Reviewer_z6Y8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem of zero-shot cross-city mobility trajectory generation, which involves generating individual trajectories for a target city. The authors assume that an individual’s daily activity schedule can be transferred as-is to the target city, requiring only the selection of new locations for each activity, and propose the SHAPE framework based on this assumption. SHAPE consists of three stages: In the first stage, a shared encoder and cross-city contrastive learning are used to align grid cells from the two cities into a common embedding space based on the functional similarity of POIs. In the second stage, a Transformer encoder, combined with a three-layer RFSQ, encodes each activity in the source city into a discrete, city-agnostic activity prototype. In the third stage, within the target city, candidate rings are delineated around the previous location based on opportunity ranking; the candidate grid cells are then scored against the prototypes to select the locations for each activity. Using three datasets and four transfer directions, SHAPE achieved the best results across all four metrics in all four directions.

Critical strengths
  1. This paper clearly defines the zero-shot cross-city mobility generation task, explicitly stipulating that the target city is represented solely by static maps and residential distribution, and the problem is clearly formulated.
  2. This paper uses discrete activity prototypes as a transfer mechanism; each prototype can be interpreted in terms of activity type, making it easier to examine than continuous latent representations.
  3. The paper uses opportunity ranking instead of physical distance to measure each step, thereby reducing reliance on the geographic scale of the source city.
Critical weaknesses
  1. The zero-shot setting is inconsistent: during Stage 1 training, a map of the target city was used, and the residential distribution was estimated from check-in data for that city.
  2. The authors state that locations are assigned based on opportunity rather than distance, yet the scoring formula includes a distance term, and the oracle experiment treats the distance distribution as additional information.
  3. The dataset is small. E.g., in the Paris test, only 20 users checked in more than 10 times, and the check-in data itself is subject to demographic bias.
  4. There are also three sets of duplicate entries in the reference list. González 2008, Jiang 2016a and 2016b, Wang 2024a and 2024b.
  5. Stage 3 has more than ten manually set hyperparameters, and the paper did not perform a sensitivity analysis.
  6. Very like an AI-generated paper.
Questions for authors

Please refer the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.