acceptodds
Under review as a conference paper at ICLR 2027

NARA: Anchor-Conditioned Representation Learning for Heterogeneous Vector Geoentities

14 views

Abstract

Vector geospatial data represent the world as discrete geoentities, such as roads, buildings, and points of interest, each with semantic attributes, geometry, and spatial relations to other geoentities, including metric proximity and topology. Existing methods for learning geoentity representations typically support a single geometry type or model only a subset of these relations, limiting their ability to capture spatial context across heterogeneous geoentities and support diverse downstream tasks. We propose NARA (Neural Anchor-conditioned Relation-Aware representation learning), a novel self-supervised representation framework for heterogeneous vector geoentities. NARA contextualizes geoentities through spatial-context-aware attention that models spatial autocorrelation using geometry distance modulated by topological relations across surrounding points, polylines, and polygons. NARA introduces masked geoentity semantic modeling and geometry-aware spatial relation modeling, as well as relation-conditioned regularization that encourages similar representations for geoentities sharing the same spatial relation to a common reference entity, while accounting for spatial autocorrelation. NARA's frozen, task-agnostic encoder outperforms state-of-the-art methods, each with an architecture tailored to its respective task, across traffic-speed prediction for polylines, building-function classification for polygons, and next point-of-interest prediction for points.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Vector geospatial representation learning, Self-supervised learning, Spatial relation learning, Heterogeneous geometric data, Spatial contextualization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_E8a8OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NARA, a self-supervised representation learning framework for heterogeneous geospatial entities. NARA introduces topology-aware spatial attention and anchor-conditioned regularization. Experiments on three downstream tasks demonstrate consistent improvements over existing baselines.

Critical strengths
  1. The method integrates semantic, geometric, and topological information in a framework.

  2. The experimental results are strong across three downstream tasks. The ablation studies further support the effectiveness of the proposed components.

Critical weaknesses
  1. The authors have clearly reduced the line spacing to include more content into main text. I am not sure if this complies with ICLR policy.

  2. The experiments section states CityFM as specifically designed for certain downstream tasks. However, CityFM is a self-supervised framework for learning geoentity representations. The claimed novelty of heterogeneous entity representation learning is also overstated.

Questions for authors
  1. See weakness.

  2. The topology-modulated distance bias consists of two additive terms, but their individual contributions are not separately validated through ablation studies.

4 comments

We sincerely appreciate the reviewer for recognizing NARA’s novelty in contextual representation learning with semantics, geometry, and their spatial relations, as well as NARA’s strong performance across tasks. We appreciate the concerns regarding presentation, our positioning relative to CityFM, and the parameterization of the attention bias. We carefully addressed each point below.

W1. The authors have clearly reduced the line spacing to include more content into main text. I am not sure if this complies with ICLR policy.

Thank you for bringing this formatting concern to our attention. Our submitted manuscript included local spacing adjustments around several section headings to satisfy the 9-page submission limit. To fully address the concern, we removed all spacing adjustments and restored the default spacing specified by the official ICLR template throughout the manuscript. We have carefully verified that the revised manuscript remains within the permitted 10-page limit.

W2. The experiments section states CityFM as specifically designed for certain downstream tasks. However, CityFM is a self-supervised framework for learning geoentity representations. The claimed novelty of heterogeneous entity representation learning is also overstated.

Thank you for this helpful clarification. We agree that CityFM is a state-of-the-art self-supervised framework for learning geoentity representations. We note that CityFM uses separate encoders for each geometry type, objective, and feature pipeline, all tailored for specific downstream tasks, whereas NARA applies a shared contextual encoder and shared self-supervised objectives across points, polylines, and polygons.

CityFM designs and uses encoders and pretraining objectives targeted to each evaluated downstream task (i.e., not task-agnostic). For example, one of CityFM’s pretraining objectives aligns an image representation of a rasterized building footprint with the semantic representation of the same building. The underlying assumption is that rasterized images help capture building shapes, and that building shapes matter for function classification. This objective would only work for downstream tasks that depend on polygon shapes (e.g., building classification), and would not work directly for points and polylines. Another example is that CityFM leverages participation in public transit routes to design contrastive learning on polyline embeddings, assuming that road polylines on shared transit routes have similar traffic speed, useful for mobility applications. Although CityFM effectively models task-relevant assumptions in pretraining objectives, design choices require redesigning pretraining objectives as new tasks emerge.

In contrast, NARA uses a task-agnostic self-supervised pretraining strategy. NARA applies the same objectives to points, polylines, and polygons rather than encoding task-specific assumptions into separate objectives. NARA learns contextual geoentity representations from complete geometry and semantics, as well as spatial relations across surrounding geoentities. This task-agnostic design is useful when a geoentity’s function depends not only on its isolated attributes or shape but also on its heterogeneous spatial context. For example, how a building polygon relates to nearby roads, POIs, and other polygons. With task-agnostic design, NARA improves traffic MAE from 3.20 to 2.93 under the matched CityFM protocol and building Macro-F1 from 69.34 to 74.30.

We view CityFM’s task-informed objectives and NARA’s task-agnostic objectives as complementary. CityFM representations can capture task-relevant information, and NARA contributes task-agnostic geoentity contextualization. Future work could further investigate ways to combine these design principles, e.g., using NARA embeddings as universal base representations and further fine-tuning with a task-relevant training objective. We have revised the manuscript to avoid claiming novelty for heterogeneous geoentity representation learning itself. We also added a statement regarding CityFM in the experiment section (Section 5; Line 357 - 360). We thank the reviewer for the opportunity to clarify this distinction and for facilitating this valuable discussion.

Q2. The topology-modulated distance bias consists of two additive terms, but their individual contributions are not separately validated through ablation studies.

Thank you for this comment. We separately ablate the topology-specific offset βtij\beta_{t_{ij}} and the topology-modulated distance term wtij⊤ϕ(dˉij)w_{t_{ij}}^\top \phi({\bar{d}_{ij}}).

Pairwise attention biasMacro-F1 ↑Weighted-F1 ↑Accuracy ↑
No pairwise bias70.92±1.491.45±0.391.97±0.3
Topology intercept only72.73±1.492.38±0.392.75±0.3
Topology-modulated distance RBF only74.82±1.492.93±0.393.31±0.2
Topology intercept + topology-modulated distance RBF (Full NARA)74.30±1.292.76±0.393.13±0.3

The topology intercept improves Macro-F1 by 1.81 points when used without the RBF term (pHolmp_{Holm} =0.0078). The topology-modulated RBF term improves Macro-F1 by 3.91 points without the intercept (pHolmp_{Holm} <0.001) and by 1.57 points when added to the topology-only model (pHolmp_{Holm} =0.021). Adding the intercept to the RBF-only model changes Macro-F1 by -0.53 points, which is not significant (pHolmp_{Holm} =0.217). This finding is consistent with our response to Reviewer H9R4 W1. Under the current distance measures (i.e., minimum distance between complete geometry), the relation-specific RBF term can already represent relation-specific values at d=0. We use both terms in the general formulation because it accommodates additional topological relations and alternative distance measures with within-relation distance variation, and because each term independently improves performance. The topology-modulated RBF term provides the largest improvement, and the full two-term formulation performs comparably to the RBF-only variant. We added the ablation results and clarified this interpretation in Appendix C.4.1 (Line 1125 - 1177).

Thanks for your detailed clarification, which addresses my concerns. I therefore raise my score.

Thank you very much for your reconsideration and for raising your score to weak accept. We are pleased that our clarifications have addressed your concerns.

Reviewer_H9R4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors develop a self-supervised framework NARA for learning contextual representations of points, polylines, and polygons. It combines semantic and geometry embeddings through a Transformer with distance- and topology-aware attention. Pre-training uses masked semantic reconstruction, spatial-relation prediction, anchor-conditioned contrastive learning (ACC), and relational semi-variogram regularization. The encoder is pre-trained separately for New York City and Singapore and kept frozen for traffic-speed estimation, building-function classification, and next-POI prediction.

Critical strengths
  1. NARA handles complete geometries together with heterogeneous spatial context, and its representations improve predictions for roads, buildings, and POIs.

  2. Based on the experiment results, the performance improvements of NARA in comparison to baselines are significant and consistent.

Critical weaknesses
  1. In Section 4.2, the authors state that CONTACT and CONTAINMENT pairs both have zero minimum geometry distance. Only DISJOINT pairs have varying distance. Under such definitions, the bias calculated by Equation 8 can be represented exactly by a shared distance function plus topology-specific offsets. Table 10 shows benefits from different parameterization, but it is not convincing enough to demonstrate the richer topology–distance interaction claimed by the paper.

  2. Based on the experiment results, it is unclear whether or not both anchor-conditioned objectives are needed across tasks. Their ablations are limited to building classification. Adding RSR improves Macro-F1 from 73.982 to 74.298 (as presented in Table 4), with almost unchanged Weighted-F1 and slightly lower accuracy. Here, paired significance is not reported. The paper also lacks a full-model comparison that removes ACC while keeping RSR.

Questions for authors
  1. Is it possible to report the full model without ACC while retaining RSR, together with paired differences across folds for the ACC and RSR comparisons?

  2. For RSR, how much of the benefit depends on distance matching and binning?

5 comments

We sincerely appreciate the reviewer’s recognition of NARA’s contextual geoentity representation learning and consistent performance on tasks covering heterogeneous geometry types, as well as the detailed technical assessments. In response to the reviewer’s comments on topology-conditioned distance bias, anchor-conditioned objectives, and RSR distance binning, we refine our claims and provide additional ablations.

W1. In Section 4.2, the authors state that CONTACT and CONTAINMENT pairs both have zero minimum geometry distance. Only DISJOINT pairs have varying distance. Under such definitions, the bias calculated by Equation 8 can be represented exactly by a shared distance function plus topology-specific offsets. Table 10 shows benefits from different parameterization, but it is not convincing enough to demonstrate the richer topology–distance interaction claimed by the paper.

Thank you for this observation. We agree that, under the current definition of distance (i.e., minimum distance of complete geometry), CONTACT and CONTAINMENT pairs occur only at d=0d=0. For either class, the RBF contribution wtij⊤ϕ(0)w_{t_{ij}}^\top \phi(0) is therefore constant across all observed pairs for specific topological relations wtijw_{t_{ij}} and is absorbed into its topology-specific offset βtij\beta_{t_{ij}}. In contrast, for DISJOINT pairs, the distance varies continuously. It distinguishes nearby but nonintersecting entities from distant ones, which topology alone cannot capture. The topology–distance interaction currently represented in NARA therefore consists of topology-specific effects at zero distance for CONTACT and CONTAINMENT, and distance-varying effects for DISJOINT pairs.

We use the current Eq. (8) as a general formulation for future work extensibility that can accommodate additional topological relations and also alternative distance measures (e.g., one that can vary for the CONTAINMENT relation). For example, all POIs contained in a mall have zero minimum geometry distance to the mall polygon, but their distances to each side of the polygon vary. Capturing such distance measures could help further characterize each POI in NARA's future work; e.g., if a POI is far from every side of the polygon, it might be far from any mall entrance. Under such definitions, CONTAINMENT pairs could have varying distance values, making relation-specific distance effects identifiable.

We agree that for the current distance measure used in NARA (minimum distances), a simpler equivalent design is a topology-specific offset plus a distance function applied only to DISJOINT pairs. We revised the manuscript to clarify the design and topology–distance interaction claim in Section 4.2 (Line 1178 - 1205). We also added a discussion of possible alternative topological relations and distance measures in Appendix C.4.1 (Line 1125 - 1177). We thank the reviewer for the opportunity to share the topology-modulated distance bias term and for facilitating this valuable discussion.

W2. Based on the experiment results, it is unclear whether or not both anchor-conditioned objectives are needed across tasks. Their ablations are limited to building classification. Adding RSR improves Macro-F1 from 73.982 to 74.298 (as presented in Table 4), with almost unchanged Weighted-F1 and slightly lower accuracy. Paired significance is not reported here. The paper also lacks a full-model comparison that removes ACC while keeping RSR.

Q1. Is it possible to report the full model without ACC while retaining RSR, together with paired differences across folds for the ACC and RSR comparisons?

Thank you for raising this point. We added an ablation that removes ACC while retaining RSR and evaluated paired differences across matched folds. This comparison is also included in the leave-one-out study reported in our response to Reviewer gbe9 W5.

On Singapore building classification, removing ACC reduces Macro-F1 from 74.30±1.2 to 72.86±1.0. The reduction is significant under a two-sided paired t-test across the ten folds after Holm correction (pHolmp_{Holm}=0.0048). Weighted-F1 and accuracy also decrease significantly, with pHolm<0.05 for both metrics. Removing RSR changes Macro-F1 from 74.30±1.2 to 73.98±1.6, which is not significant (pHolmp_{Holm}=1.000). Weighted-F1 and accuracy remain effectively unchanged.

We also conduct the corresponding ablation on NYC traffic speed estimation. Removing RSR increases MAE from 2.93 to 2.95, with a paired effect of 0.0172 (pHolmp_{Holm}=0.0213). Removing ACC changes MAE by only 0.0012, with no significant effect (pHolmp_{Holm}=0.9133).

The table below summarizes the paired effects of ACC and RSR under two-sided paired t-tests. The effect is full NARA minus the corresponding ablated model. Positive effects favor full NARA for Macro-F1, while negative effects favor full NARA for MAE. Holm correction is applied within each task.

Task and metricComponentFull NARAWithout componentPaired effectpHolmp_{Holm}
SG Building Macro-F1 ↑ACC74.30 ± 1.272.86 ± 1.0+1.440.0048
SG Building Macro-F1 ↑RSR74.30 ± 1.273.98 ± 1.6+0.321.000
NYC Traffic MAE ↓ACC2.93 ± 0.012.93 ± 0.01−0.00120.9133
NYC Traffic MAE ↓RSR2.93 ± 0.012.95 ± 0.01−0.01720.0213

ACC and RSR capture anchor relational context. ACC aligns geoentities that share an anchor and relation, while RSR regulates their representation dispersion relative to nonsibling pairs. The results reflect these distinct roles. ACC significantly improves building classification, while RSR significantly improves traffic estimation. On the task where an objective has no significant effect, the full model does not incur a meaningful performance degradation and remains above the state-of-the-art baseline, CityFM. We added the ACC/RSR ablation results and discussion in Appendix C.4.2 (Line 1178 - 1205).

Q2. For RSR, how much of the benefit depends on distance matching and binning?

Thank you for this question. We conducted a 2 x 2 ablation, varying the RSR distance cutoff (50 or 100 m) and the number of distance bins (5 or 10). The 50 m/5-bin configuration used in the paper performs best on both tasks, achieving 2.93 ± 0.01 MAE on NYC traffic and 74.30 ± 1.2 Macro-F1 on SG building. The maximum lag (i.e., 50 m, distance cutoff) determines the spatial extent being summarized, while lag width (i.e., 10 m) trades finer spatial resolution against the number and stability of pairs within each bin. The current setting works well for denser, large cities, where the extent of spatial autocorrelation happens at a smaller distance (e.g., 50 m or half a street block) and within a local neighborhood (e.g., a storefront would be 5-10 meters).

(a) Traffic speed prediction (NYC)

Distance CutoffBinsRMSE ↓MAE ↓R² ↑MAPE (%) ↓
50 m53.81 ± 0.012.93 ± 0.010.80 ± 0.0017.68 ± 0.1
50 m103.84 ± 0.022.97 ± 0.010.80 ± 0.0017.92 ± 0.2
100 m53.84 ± 0.022.96 ± 0.010.80 ± 0.0017.78 ± 0.1
100 m103.81 ± 0.012.94 ± 0.010.80 ± 0.0017.75 ± 0.1

(b) Building function classification (SG)

Distance CutoffBinsMacro-F1 ↑Weighted-F1 ↑Accuracy ↑
50 m574.30 ± 1.292.76 ± 0.393.13 ± 0.3
50 m1072.81 ± 1.792.58 ± 0.492.99 ± 0.3
100 m572.70 ± 1.492.41 ± 0.492.82 ± 0.3
100 m1073.42 ± 1.692.62 ± 0.493.06 ± 0.3

With five bins, extending the cutoff from 50 to 100 m increases NYC MAE from 2.93±0.01 to 2.96±0.01 and reduces Singapore Macro-F1 from 74.30±1.2 to 72.70±1.4. At 50 m, increasing from five to ten bins also degrades both metrics (2.97±0.01 MAE and 72.81±1.7 Macro-F1). At 100 m, ten bins partially recover performance (2.94±0.01 MAE and 73.42±1.6 Macro-F1) but remain below the selected 50-m/five-bin configuration.

Notably, 50 m/five bins and 100 m/ten bins both use 10-m intervals. Their difference therefore isolates the matching cutoff rather than bin width, indicating that RSR benefits more from short-range relational context than from extending matching to 100 m. In contrast, the comparison between five and ten bins at a fixed cutoff shows that finer resolution does not consistently improve performance. The selected configuration balances a local 50-m context with sufficient pair aggregation over five 10-m intervals. We added ablation results and discussion for this binning and max-distance limit in Appendix C.4.3 (Line 1207 - 1231).

Reviewer_gbe9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes NARA, a novel self-supervised framework for heterogeneous geoentities (points, polylines, and polygons) representation. The critical designs are self-supervised objectives, which are geographic entity mask reconstruction, spatial relation modeling, and relation-conditioned regularization. Experiments show that NARA outperforms baselines on three urban tasks.

Critical strengths

The paper's strength lies in combining heterogeneous geoentities, such as Point/Polyline/Polygon, to perform geographic representation learning.

Critical weaknesses
  1. Semantic dimension and geometric dimension seem to be different. Is there a way to align them?

  2. It is not clear what the impact of the number of removed false negatives is in Eq. (9).

  3. How to evaluate the importance among the three losses in Eq. (25) after their numerical scales are matched?

  4. This paper lacks a comparison of training/inference costs between the baselines and the proposed methods.

  5. The leave-one-out strategy in the ablation study can evaluate the contribution of each component. I recommend that the authors present corresponding experimental results.

Questions for authors

See the Weaknesses

4 comments

We sincerely thank the reviewer for thoughtful comments and for recognizing the value of learning contextual representations of heterogeneous geoentities. The comments helped us clarify semantic–geometry fusion and false-negative filtering and motivated additional ablations. We respond to each question below.

W1. Semantic dimension and geometric dimension seem to be different. Is there a way to align them?

Yes, the semantic and geometric dimensions differ at the encoder outputs. The semantic encoder generates a normalized 768-dimensional representation, while Poly2Vec generates a 128-dimensional geometric representation. To align them, NARA uses separate learned projection heads to map both to the shared dimension dmodeld_{model} =128, yielding zisemz_{i}^{sem} and zigeomz_{i}^{geom} (Section 4.1 and Appendix B.1), and then uses an adaptive gate to fuse them (in Eq. (5)). The projection heads and gate are jointly learned with NARA’s self-supervised pretraining objectives.

W2. It is not clear what the impact of the number of removed false negatives is in Eq. (9).

Thank you for requesting this clarification. We quantified how frequently geoentities with identical semantic embeddings occur within pretraining spatial windows and conducted an ablation that retains these candidates in the MGSM negative set.

At least one identical semantic pair occurs in 66.90% of Singapore windows and 72.84% of NYC windows. Among queries eligible for masking, 87.29% in Singapore and 77.00% in NYC have at least one identical semantic candidate excluded. For each query geoentity (i.e., the target masked geoentity), these two candidates cover, on average, 22.27% and 28.15% of the negative geoentity pool, respectively.

MGSM negative filteringMacro-F1 ↑Weighted-F1 ↑Accuracy ↑
Filter identical semantics candidates74.30 ± 1.292.76 ± 0.393.13 ± 0.3
Retain identical semantics candidates69.92 ± 1.391.27 ± 0.391.83 ± 0.2
Difference (retain − filter)−4.38−1.49−1.30

These results demonstrate that false negatives in the contrastive negative set substantially degrade NARA’s representation quality. All paired differences are significant under Holm-corrected two-sided paired t-tests (pHolmp_{Holm}< 0.001). Without filtering, MGSM pulls the prediction toward the target semantic embedding while simultaneously pushing it away from an identical semantic embedding. Eq. (9) excludes the corresponding query-candidate pairs only from the denominator of the MGSM contrastive objective, while all geoentities remain in the input spatial window. We report the diagnostic result in Appendix C.4.4 (Line 1233 - 1254).

W3. How to evaluate the importance among the three losses in Eq. (25) after their numerical scales are matched?

Thank you for raising this question. Following the reviewer’s suggestion in W5, we evaluate the three auxiliary objectives, GSRM, ACC, and RSR, by removing one objective at a time and retraining NARA from scratch while retaining all other final settings. The complete numerical results and paired significance tests are reported in our response to W5 below. In summary, GSRM and ACC make the largest contributions to Singapore building classification. RSR has a smaller effect on the SG building task but significantly improves the NYC traffic task, showing that the relative contribution of RSR/ACC depends on the downstream task. We report these leave-one-out ablation results in Appendix C.4.2 (Line 1178 - 1205). Appendix C.4.5 (Line 1256 - 1275) additionally provides a sensitivity ablation in which we vary each loss ratio one at a time.

W4. This paper lacks a comparison of training/inference costs between the baselines and the proposed methods.

Thank you for this valuable comment. We benchmark CityFM and NARA on a single NVIDIA RTX A5000.

ModelParametersSG training unit (min)NYC training unit (min)SG inference (ms/entity)NYC inference (ms/entity)
CityFM123.71M6.795.6110.8911.58
NARA1.71M2.404.2514.6417.41

NARA requires 64.7% and 24.2% less time per reported training unit for Singapore and NYC, respectively, and has about 72 times fewer trainable parameters than CityFM. The training units follow each model’s released procedure and are not equivalent to optimization epochs.

For CityFM, one training unit includes one epoch in each of its three sequential training stages. For NARA, one training unit is one pretraining epoch. CityFM does not report the number of training units required for convergence, so we do not estimate its total training cost.

For inference, CityFM requires 10.89 ms per Singapore building and 11.58 ms per NYC road. NARA requires 14.64 and 17.41 ms, respectively. NARA’s extraction procedure queries and encodes the semantics and complete geometry of nearby entities for each target. CityFM uses neighboring semantic representations and engineered features. Both methods extract representations once and cache them for downstream evaluation. We report these measurements and implementation details in Appendix B.5 (Line 1004 - 1025).

W5. The leave-one-out strategy in the ablation study can evaluate the contribution of each component. I recommend that the authors present corresponding experimental results.

We thank the reviewer for the suggestion. We conduct a leave-one-out ablation from the full NARA, removing one component at a time.

ConfigurationMacro-F1 ↑Weighted-F1 ↑Accuracy ↑Δ Macro-F1
Full NARA74.30 ± 1.292.76 ± 0.393.13 ± 0.3---
− topology-modulated distance bias70.92 ± 1.491.45 ± 0.391.97 ± 0.3−3.38
− GSRM72.81 ± 1.792.33 ± 0.492.75 ± 0.3−1.48
− ACC72.86 ± 1.092.39 ± 0.292.82 ± 0.2−1.44
− RSR73.98 ± 1.692.76 ± 0.393.14 ± 0.3−0.32

Removing the topology-modulated distance bias causes the largest reduction in Macro-F1 at 3.38 points. Removing GSRM and ACC reduces Macro-F1 by 1.48 and 1.44 points, respectively. Removing RSR changes Macro-F1 by 0.32 points and does not materially change Weighted-F1 or accuracy. The paired Macro-F1 difference for RSR is not significant across the ten folds (pHolmp_{Holm} = 1.000). These results indicate that GSRM and ACC contribute more to SG building task, while RSR retains comparable performance on the SG building task.

We further evaluate ACC and RSR on NYC traffic estimation to examine whether their contributions differ across downstream tasks (based on Reviewer H9R4 W2 suggestion). The table below reports two-sided paired t-tests over matched folds or seeds. The paired effect is computed as full NARA minus the corresponding ablated NARA.

Task and metricComponentFull NARAWithout componentPaired effectpHolmp_{Holm}
SG Building Macro-F1 ↑ACC74.30 ± 1.272.86 ± 1.0+1.440.0048
SG Building Macro-F1 ↑RSR74.30 ± 1.273.98 ± 1.6+0.321.000
NYC Traffic MAE ↓ACC2.93 ± 0.012.93 ± 0.01−0.00120.9133
NYC Traffic MAE ↓RSR2.93 ± 0.012.95 ± 0.01−0.01720.0213

RSR significantly reduces NYC traffic MAE by 0.0172 (pHolmp_{Holm} =0.0213), while its effect on Singapore building is modest and not significant. The full model therefore preserves building classification performance while obtaining a significant improvement in the NYC traffic task. ACC shows the complementary pattern. It significantly improves building Macro-F1 by 1.44 points (pHolmp_{Holm} =0.0048) but has no detectable effect on traffic MAE (pHolmp_{Holm} =0.9133). We report the complete leave-one-out ablation and paired comparisons in Appendix C.4.2 (Line 1178 - 1205).

Thanks for the response. I have no further questions and recommend acceptance after the weaknesses are revised.

Thank you very much for your positive feedback and recommendation for acceptance.

We have also uploaded the revised manuscript, with all changes from the initial submission highlighted in blue for ease of reference.

If any specific weaknesses still require revision, we would greatly appreciate your clarification so that we can address them accordingly.

Thank you again for your valuable comments and support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.