acceptodds
Withdrawn from ICLR 2027

BAViR: Aligning Urban Visual Representations with Revealed Mobility Behavior

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Urban environment representations should describe how places are used, not only their visible structure. Mobility records reveal this use but need not be available wherever a representation is deployed. We use mobility behavior as training supervision to learn an environmental representation that can be applied where mobility observation are unavailable. We introduce BAViR (Behavior-Aligned Visual Representations), which fuses satellite and street imagery with points of interest and road structure, and a contrastive objective aligns its output with visitation, dwell, activity, and temporal summaries. Since mobility behavior is used for supervision only during training, the learned encoder can represent new locations from environmental observations alone. We evaluate spatial generalization and reuse for targets excluded from encoder training, using two five-fold spatial protocols and three seeds in metropolitan Orlando. On geographically withheld regions, BAViR improves visit-count from 0.488 to 0.523 over matched masked-stream self-supervision. Moreover, the learned representation also transfers to transaction spend, a target excluded from encoder training, reaching of 0.355 compared with 0.314 for self-supervision and 0.339 for behavior regression under the same frozen-feature readout. With only 10% of downstream spend labels, BAViR exceeds raw-feature Ridge trained with all labels. Together, these results support the value of behavior supervision for learning reusable functional representations, rather than a universal advantage of contrastive learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
BAViR, urban representation learning, multimodal fusion, contrastive learning, self-supervised learning, human mobility, geospatial machine learning, spatial cross-validation, remote sensing

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_dCV2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes a cross-modal representation alignment approach, inspired by CLIP’s image–text alignment, to incorporate mobility information into urban representations learned from satellite imagery, street-view imagery, POIs, and road network structure. Experiments in the Orlando metropolitan area show that this approach can successfully enrich these representations with information about human activity.

Critical strengths
  1. The paper is well motivated, focusing on the reusability of urban representations, an important goal that many existing studies struggle to achieve, often without adequately discussing this limitation.
  2. The intuition behind the methodological choice is sound. CLIP-style cross-modal alignment is a natural and straightforward approach to achieving the paper’s objective.
Critical weaknesses
  1. The writing is often difficult to follow, as defensive remarks frequently interrupt otherwise coherent arguments. These interruptions disrupt the logical flow and detract from the clarity of the presentation.
  2. Although the methodological intuition is sound, the paper overlooks an important consideration: this learning paradigm typically requires very large-scale training data to achieve its intended goals. However, the experimental evaluation lacks scaling studies that examine how the approach benefits from increasing amounts of data, leaving this important aspect unexplored.
  3. Urban representation learning has been actively studied since 2016 and has accumulated an extensive body of work. But this paper lacks adequate experimental comparisons with, and discussion of, these works. Instead, the evaluation remains largely confined to its own framing, making it difficult to position the contribution relative to existing work or identify the broader insights it offers to the field.
Questions for authors

Please refer to weakness.

Reviewer_yS4qOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces BAViR, a representation learning framework that fuse satellite imagery, street imagery, POI and road structure with mobility behaviour, unders contrastive objective. Since mobility is used only for supervision at the training time, the learned encoder can be applied at locations where only environmental observations are available. The authors further test the scenario where limited downstream labels are used. Experiments in Orlando demonstrate improvements and the transfer to unseen downstream targets, particularly transaction spending.

Critical strengths
  • The authors use two spatial holdout protocols (interleaved and geo-block) and group cells by H3 parent units. This keeps spatial autocorrelation from leaking information between training and test sets, which in important for spatially smooth target prediction. The geo-block setting in particular is close to deploying the model in an unobserved region.
  • The authors acknowledge prior work on mobility-free deployment and frame the paper as a controlled study of the value of mobility supervision under different training objectives instead of a new SOTA method.
Critical weaknesses
  • While controlled baselines serve as a highlight for this paper, their implementations are not sufficiently described. For example, how do you construct positive/negative pair for Environment InfoNCE? These details are important for assessing the fairness and reproducibility of the comparisons.
  • The benefit of BAViR is most pronounced in card-spend prediction (which could be correlated with visit counts), while Table 10 reports that raw-feature regression performs better on the physical and crime targets. This suggests that the transfer benefit of mobility alignment may be limited in scope.
  • The writing is not well organized, and some important information is buried or missing, which makes the paper harder to follow. For example, line 266 describes the spatial holdout procedure using terms such as "H3 parent group" for the first time, without explaining them or pointing readers to the details in the Appendix.
  • The methodology is standard, and the lack of empirical comparison with existing mobility-free deployment methods, together with validation on only one city, limits the contribution.
Questions for authors
  • Could the authors provide more detailed descriptions of the controlled baselines?
  • Why does BAViR perform worse on crime prediction, which is also closely related to human activity?
  • What is the central contribution of BAViR, given existing work on both the value of mobility data and mobility-free deployment? Could the authors provide further justification?
Reviewer_iWTSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BAViR, a framework for learning transferable urban environment representations by leveraging mobility behavior as a training-time supervision signal. The central idea is to align visual or environmental features with human mobility patterns during representation learning, enabling the resulting representations to be reused for downstream tasks without requiring mobility data at inference time.

The authors evaluate the learned representations on downstream tasks such as card-spending prediction, including spatial holdout settings in which the target locations are excluded from encoder training. The paper also investigates whether behavior-supervised representations capture useful functional information beyond the mobility statistics used during training.

The experimental evaluation includes multiple baselines and controlled comparisons, with additional analyses highlighting the strengths and limitations of the proposed representation.

Overall, the paper addresses an interesting and potentially useful research problem, and the experimental evaluation appears reasonably comprehensive. However, I find the methodological contribution rather limited, as the proposed framework seems to rely primarily on a relatively straightforward encoder-and-Transformer architecture without introducing substantial technical innovations.

Critical strengths
  1. An interesting research problem. The paper investigates an interesting direction: leveraging mobility behavior to supervise the learning of visual and environmental representations. Learning representations that capture the functional characteristics of urban environments, rather than relying solely on their visual appearance, is a meaningful research objective with potential applications in urban computing.

  2. Extensive experimental evaluation. The authors compare the proposed approach against a diverse set of baselines and conduct extensive experiments to evaluate its effectiveness. The inclusion of multiple controlled comparisons and the reporting of variance across experimental runs further strengthen the empirical evaluation.

  3. A relatively complete study. The paper presents a coherent research pipeline, from problem formulation and methodological design to experimental evaluation and discussion of limitations. Overall, the work appears reasonably complete and provides empirical evidence for the potential benefits of behavior-supervised representation learning.

Critical weaknesses
  1. Limited methodological novelty and technical depth. My primary concern is the lack of substantial technical innovation in the proposed framework. As I understand it, the method mainly relies on conventional encoders and Transformer-based feature processing to align environmental representations with mobility-derived supervision signals. While this design may be effective, the overall architecture appears relatively straightforward and does not seem to introduce a particularly novel modeling mechanism or learning paradigm. The paper's contribution therefore appears to lie primarily in the problem formulation and application of existing techniques rather than in methodological advances. For a venue such as ICLR, I would expect a stronger technical contribution or a more convincing demonstration of why the proposed design goes beyond a straightforward combination of established components.

  2. Suboptimal figure quality. The figures appear to be embedded as raster images (e.g., PNG) rather than vector graphics. This may compromise their visual clarity, particularly when the manuscript is magnified or printed. For a top-tier conference submission, figures containing diagrams, text, and plots should preferably be provided in vector formats such as PDF or SVG whenever possible. This is a relatively basic aspect of academic presentation that deserves more careful attention.

  3. Inefficient use of space and poor manuscript layout. The overall manuscript layout is unnecessarily sparse. For example, Figure 4 occupies a disproportionately large amount of space, and similar issues can be observed in other figures. This results in substantial unused whitespace and disrupts the reading flow. In particular, the region around Lines 425–471 appears to contain excessive blank space that could be avoided through more careful formatting. A more compact figure arrangement and improved page layout would substantially enhance the presentation quality and information density of the paper.

Questions for authors
  1. On methodological novelty: Could the authors clarify which components of BAViR constitute the main technical innovations beyond conventional encoder-based representation learning and Transformer architectures? In particular, what distinguishes the proposed approach from a straightforward multimodal alignment framework using mobility-derived supervision?

  2. On the necessity of the proposed architecture: Have the authors compared BAViR with simpler alternatives, such as standard visual encoders followed by an MLP projection head or conventional multimodal contrastive learning frameworks? Such comparisons would help establish whether the proposed architectural choices are necessary for the reported improvements.

  3. On figure quality: Could the authors improve the figures by using vector-based formats where appropriate and ensuring that all text and graphical elements remain clear when magnified?

  4. On manuscript layout: Could the authors revise the layout of Figure 4 and other oversized figures to reduce unnecessary whitespace, particularly around Lines 425–471?

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

Related papers

Discussion (0)

Sign in to comment.