- 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
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.
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.
Detailed responses:
(1) Effect of pretraining.
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.
Edges.
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.
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.