acceptodds
Under review as a conference paper at ICLR 2027

ArcX: A Unified Relation Encoder for Vector Spatial Representation Learning

16 views

Abstract

Vector spatial representation learning (VSRL) has primarily focused on representing individual spatial entities, such as their locations, geometries, and attributes, while paying limited attention to pairwise spatial entity relations. Current methods face two limitations in spatial relation representation. First, they often rely on centroid distance, which can discard geometric details. Second, relations are typically inferred only after the two entities are encoded independently, resulting in limited spatial interaction. We introduce pair-level VSRL as a complementary representation perspective and propose ArcX, a unified VSRL encoder that generates dedicated spatial relation embeddings for entity pairs. ArcX introduces an arc-wise cross-entity modulation mechanism, allowing each entity to respond to the other before embedding readout. We evaluate ArcX across five datasets spanning pair-, entity-, and scene-level VSRL tasks. The results show that ArcX can capture multiple dimensions of spatial relations, outperforming the second-best method by an average of 10.8% on pairwise tasks. ArcX can also improve entity-level neighborhood aggregation by an average of 4.4% over the second-best method and scene-level information propagation by an average of 16.5%, demonstrating consistent utility across diverse task settings. ArcX introduces a relation-centric perspective to VSRL, complementing existing entity- and scene-level perspectives. The code and data are available at https://anonymous.4open.science/r/ArcX-6BB3.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
spatial relation, spatial representation learning, graph neural network, vector data

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_a4CdOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a relation-centric encoder for vector spatial representation learning. It focuses on pair-level spatial relations between heterogeneous geospatial entities. The proposed ArcX model tokenizes entities into arcs, performs arc-wise cross-entity modulation in a pair-local coordinate system, and reads out a relation embedding by jointly capturing distance, direction, and topology. This is then used to modulate message passing in downstream entity- and scene-level tasks. In addition, a Triple-Distance Spatial Kernel learns a mixture over minimum, centroid, and maximum distances to provide scalar affinities.

Critical strengths

For impact, the relation-centric perspective proposed by the authors fills a gap between entity- and scene-level VSRL, with potential practical benefits for a range of downstream spatial tasks. In addition, the approach is geometry-type agnostic (points, polylines, polygons), increasing applicability in real-world GIS/GeoAI systems.

As methodological innovation, the paper reframes VSRL from an entity-centric to a relation-centric perspective and proposes a dedicated pair-level encoder that produces relation embeddings, which is underexplored in prior work. Arc-wise cross-entity modulation is a simple mechanism allowing intra-entity geometric features to respond to the other entity before readout, avoiding O(n_i n_j) cross-attention. In addition, the TDSK component is a learnable, task-adaptive kernel over complementary geometric distances (min, centroid, max), which is quite interpretable.

For experiments, the paper includes fair evaluation across five datasets and three task levels (pairwise, entity-level, and scene-level) demonstrates both effectiveness and generality. Ablations are present and disentangle contributions of TDSK, ArcX, and pretraining, and show complementary effects of modeling multiple relation aspects. The paper shows consistent improvements over baselines, including attention and GNN methods, with sizable margins in several settings.

For clarity, the definitions and equations are OK; the decomposition into pair-local transform, arc-token encoding, cross-entity gating, and readout is easy to follow

Critical weaknesses

There are some technical limitations or concerns:

  • From a methodological point of view, the main and important limit is about scalability. ArcSet uses self-attention over arc tokens, which is quadratic in the number of arcs per entity; for highly detailed geometries or dense multi-polylines, computational and memory costs could be significant. The paper does not provide complexity/runtime analysis or strategies (e.g., arc subsampling) to address this.

In addition, the pair-local transformation uses translation only (no rotation/scale normalization), making the relation embedding sensitive to global coordinate orientation. The implications for rotation-variant versus rotation-invariant tasks are not discussed. It’s a real shame for a high-level conference of this kind.

  • From an experimental point of view, it lacks baseline comparisons to methods with learned edge attributes/filters, such as edge-conditioned convolutions (dynamic edge-conditioned filters) or recent adaptive edge-weight graph models, which are conceptually close to relation-conditioned message passing. Also, relation-aware pretraining benefits ArcX; so it is unclear to me whether baselines were allowed an equivalent multi-task or relation-prediction pretraining protocol to control for pretraining advantages. No runtime/memory or scalability results are reported by authors, which are important for practicality at large scales.

  • For clarity, the details on neighbor selection during message passing and how TDSK is computed efficiently are limited. Some implementation specifics are missing or scattered, making reproducibility and deployment considerations harder to assess.

  • For positioning regarding the SOA, the paper could better contextualize against relation-aware or edge-attribute learning frameworks. Recent geospatial relation-aware pretraining methods that jointly encode topology and distance (e.g., anchor-conditioned relation-aware models) are not discussed, despite conceptual overlap.

Questions for authors

How does ArcX’s runtime and memory scale with the number of arcs per entity and neighbors per node? 

Could you compare against dynamic edge-conditioned filtering baselines where edge features are learned functions of geometry, to better position ArcX’s relation embeddings as “learned” edge attributes?

3 comments
AuthorsOpenReview ↗

Response to Reviewer a4Cd

Thank you for the suggestions. We hope our response addresses your concerns and would appreciate your reconsideration in light of the evidence below.

  • Without pretraining, ArcX still outperforms all evaluated baselines, including edge-conditioned convolution.
  • Runtime and peak GPU memory grew sublinearly with arc count.
  • ArcX's contribution is different from recent spatial relation works.
  • ArcX allows flexible invariance choices.

Detailed responses:

(1) Effect of pretraining.

  1. Without pretraining, ArcX still outperforms all baselines on all tasks.
  2. On every task except Building Pattern, ArcX's architecture contributes more than pretraining to its gains over the strongest baseline.

Please see our first response to Reviewer 5zsE for full results.

(2) Why translation and when is invariance useful?

Why translation? Translation can reduce the influence of global location on local relations (Chu & Shahabi, 2026; Mai et al., 2023).

Why not enforce rotation and scale invariance? Direction and scale/distance are within the three main spatial relation types and can be central to certain tasks. Examples include direction estimation in auto-driving, wind direction analysis in meteorology, and layout requirements in urban planning. Enforcing rotation/scale invariance may make ArcX ineffective in such tasks.

When is invariance desirable? We appreciate your point that invariance is beneficial for some tasks, such as shape tasks. ArcX can still learn these invariances through data augmentation and pretraining.

(3) Recent work on relation-aware spatial representation learning.

Thank you for this point. We were aware of several recent works when preparing the manuscript, including NARA (Kim et al., 2026) and UrbanMFM (Zhang et al., 2026). NARA uses distance and topology to learn contextualized representations. UrbanMFM uses contrastive learning to capture relationships. ArcX shifts the focus from using spatial relations to help represent entities to directly representing the spatial relations themselves.

We will cite both works and clarify how ArcX relates to them in the revised manuscript. NARA UrbanMFM

(4) Neighbor selection and TDSK computation.

We use a symmetric centroid-distance-KNN graph with k=4 for Building Pattern and k=8 for other tasks. This simple construction ensures that all methods use the same neighbors, eliminating the effect of differences in graph structure on evaluation.

All TDSK distances are precomputed only once and require no computation at inference time. We use the built-in distance function in Shapely for minimum distances and rotating caliper algorithm for maximum distances. At inference, TDSK reads the cached distances directly.

(5) Runtime and memory scaling.

In ArcSet, arcs can be processed by three set models, DeepSets, PointNet, and Set Transformer. DeepSets and PointNet have O(n) complexity without expensive self-attention. The time complexity of standard self-attention is O(n²). In practice, the induced strategy recommended by Set Transformer authors is used. Its complexity is O(mn), where m is a fixed number of inducing points (Lee et al., 2019). We set m to 16.

We ran efficiency tests using an H200 GPU on the largest dataset LCZ. Each entry reports inference time (ms) / peak GPU memory (MB).

Arcs/entity.

Arcs/entityArcSetArcX
165.57/45.56.14/46.5
325.56/57.56.29/59.5
645.57/81.66.30/85.6
1285.44/129.66.33/137.6
2565.95/225.96.48/241.9

Edges.

kEdgesRelation encodingPropagation
21,37043.5/314.70.40/59.4
42,70884.1/588.90.39/84.2
85,840182.8/1,232.60.40/142.4
1611,504357.3/2,391.90.40/248.3

In theory, the cost of self-attention grows linearly with arc count, rather than quadratically.

In experiments, both runtime and peak memory scaled sublinearly when arc count or edge count increased.

(6) Comparison with edge-conditioned convolution.

We compare ECC and ArcX. Both methods are trained without pretraining and use the same graphs and TDSK configuration.

TaskMethodAccuracyParamsTraining time/step (ms)GPU memory/step (MB)
Flow DirectionECC0.6221±0.009027520911.18142.5
ArcX0.6502±0.019230839410.3279.8
LCZECC0.7011±0.0047109063611.33374.2
ArcX0.7088±0.005434564510.55114.3
Building PatternECC0.5482±0.0066109089611.1594.0
ArcX0.5692±0.007434590510.8985.9

Accuracy. Under the same conditions, ArcX achieves higher accuracy on all tasks. Since the LCZ accuracies are close, we also measured Macro-F1 (0.6580±0.0284 for ArcX; 0.6072±0.0065 for ECC). ArcX performs better when all classes are weighted equally.

Efficiency. ArcX takes less time and less GPU memory on all tasks. It has fewer params on LCZ and Building Pattern and slightly more on Flow Direction.

I thank the authors for the clarity of their responses and the additional information provided. This allows me to better appreciate certain points, but it does not lead me to change my rating, which I consider fair given the overall criticisms raised regarding the submitted manuscript.

Thank you for your response. We appreciate your time and consideration. We are glad to hear that our responses helped you better understand our work. Please feel free to leave a comment if you have any further concerns.

Reviewer_SMk6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ArcX, a unified relation-centric encoder to tackle pair-level vector spatial representation learning, aiming to bridge the gap between entity-level and scene level representation. The author identifies two major limitations of existing methods: (1) using centroid distance to describe spatial affinity discards rich spatial and geometric information, and (2) cross-entity interaction after the encoding stage is insufficient since their representations have been compressed. To address these, ArcX first uses TDSK to capture complementary geometric cues for spatial affinity. Then it uses pair-local coordinates to remove absolute-position effects while preserving relative configuration; arc-token encoding to represent heterogeneous geometries at fine-grained arc-level resolution; arc-wise cross-entity modulation to condition each arc token on the other entity before readout; and relation readout to fuse intra- and cross-entity responses into reusable relation embeddings. Finally, relation-conditioned message passing uses these embeddings to gate exchanged messages and TDSK to weight their propagation. Experiments across five datasets show that ArcX achieves the best results on pairwise topology, direction, and distance tasks, entity-level regression, and scene-level classification.

Critical strengths
  1. The pair-level design. ArcX proposes pair-level VSRL between entity-level and scene-level scope. It lets each entity’s arc features respond to the other entity before they are pooled, producing a directed relation embedding. This addresses a real limitation of combining independently encoded geometries only after readout and is the paper’s strongest methodological contribution.
  2. Verification of reusable relation embedding. The empirical evidence is broad: pairwise topology/direction/distance on Beijing and zero-shot Xi’an, entity-level BIKE/HUMOB regression, and scene-level HydroSHEDS/LCZ/Building Pattern classification. The cross-level consistency and zero-shot result substantially strengthen the case that ArcX captures transferable spatial relations rather than overfitting one task.
Critical weaknesses
  1. The pairwise comparison does not establish an advantage over the strongest recent geometry encoder. Geo2Vec[1] is discussed and used for downstream initialization, but is absent from the pairwise results. Since it also handles points, lines, and polygons and evaluates spatial relations, this omission limits the evidence for ArcX’s claimed advantage over existing encoders.
  2. The experimental results do not fully isolate the proposed interaction mechanism. ArcX improves substantially over ArcSet under the same pairwise training tasks, which is meaningful evidence. However, ArcX also adds a relation-specific branch and parameters. The paper does not compare its pooled-summary arc gate against parameter-matched alternatives such as FiLM-style[2] conditioning, bilinear interaction, or arc-level cross-attention. Consequently, the results establish the value of adding pre-readout interaction more clearly than they establish the specific value of ArcX’s modulation design.
  3. The downstream gains are not cleanly attributed to relation-conditioned propagation. ArcX receives separate distance-and-direction pretraining on each downstream dataset, while the competing aggregation methods do not receive comparable relation supervision. On Building Fragments, ArcX accuracy rises from 56.92% without this pretraining to 70.36% with it. Without matched controls across the scene tasks, it is difficult to determine how much of the reported gain comes from the propagation design.

[1] Chu, Chen, and Cyrus Shahabi. “Geo2Vec: Shape- and Distance-Aware Neural Representation of Geospatial Entities.” arXiv, 2025, arXiv:2508.19305, https://arxiv.org/abs/2508.19305. [2] Perez, Ethan, et al. “FiLM: Visual Reasoning with a General Conditioning Layer.” arXiv, 2017, arXiv:1709.07871, https://arxiv.org/abs/1709.07871.

Questions for authors
  1. Advantage against Geo2Vec: Can you evaluate Geo2Vec with the same settings used for the pairwise baselines? Would ArcX retain its advantage?
  2. Architectural comparison: Under the same pairwise supervision and a matched parameter budget, how does ArcX compare with a pooled-entity FiLM gate, a bilinear pair encoder, or arc-level cross-attention? Which part of the gain over ArcSet specifically requires the proposed arc-wise modulation?
  3. Pretraining attribution: Were any downstream aggregation baselines given equivalent distance-and-direction pretraining? If not, could you provide a matched-supervision comparison and report results on additional entity- and scene-level tasks, rather than only on Building Fragments?
2 comments

I thank the authors for providing additional results and analysis, which partially address my main concerns. However, the other reviewer’s comments highlight a conceptual similarity between the proposed approach and prior work in how they represent spatial relations. I remain unconvinced that the distinction is substantial enough to support the novelty claim, so I am maintaining my original rating.

AuthorsOpenReview ↗

Response to Reviewer SMk6

Thank you for your follow-up. We would like to clarify an important distinction in what the learned embeddings represent.

Previous studies use entity B to answer “What is entity A like?” and entity A to answer “What is entity B like?” The resulting embedding is still an embedding of A or an embedding of B.

We instead directly answer “What is the relationship between A and B like?” The embedding belongs to the spatial relationship between A and B itself, rather than to either individual entity.

Our contribution therefore concerns a change in the representation target—from using spatial relations to characterize individual entities to representing the spatial relations themselves. We believe this distinction is important when assessing the conceptual similarity between the approaches.

We also compared Geo2Vec and ArcX under a transductive protocol, evaluating Geo2Vec with two feature widths. ArcX achieves better results in seven of the eight comparisons across the two Geo2Vec variants. The only exception is Distance-MAE.

ModelTopology-Macro-F1 ↑Direction-Acc ↑Direction-MAE ↓Distance-MAE ↓
Geo2Vec-1280.73680.77605.39332.82
Geo2Vec-2560.72170.80974.92436.90
ArcX0.84680.88872.71334.43

Again, we sincerely appreciate your time and comments, which have helped us refine our ideas and strengthen this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.