acceptodds
Under review as a conference paper at ICLR 2027

Orientation-Augmented Route Topological Representation for Trajectory Clustering

Abstract

For traffic behavior analysis, trajectory clustering requires representations that make raw movement sequences comparable while suppressing irrelevant geometric variation and retaining the spatial structure needed to distinguish travel routes. Topological trajectory representations describe how paths wind around obstacles or landmarks, but flattened harmonic embeddings may not distinguish routes with similar cumulative responses and different landmark-relative orientations. We propose Azimuth-Resolved Harmonic Representation (ARHR), which supplements the cumulative harmonic embedding with a compact summary of landmark-relative angular activity. Conditional analysis gives sufficient conditions for this angular summary to distinguish trajectories with equal cumulative harmonic responses. Experiments on real-world traffic data show dataset-dependent improvements over the evaluated flattened-harmonic baseline. Weighting controls and order ablations show that clustering performance depends on the activity definition, while higher angular orders do not consistently improve clustering across datasets. These findings clarify when angular activity helps harmonic trajectory representations and support retaining this complementary structure according to the route distinctions required by the task.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
trajectory clustering, topological representation, harmonic embedding, azimuth, route discrimination

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_RbzyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ARHR, a trajectory representation for route clustering. ARHR adds a few angular statistics, describing where a trajectory moves around a chosen landmark, to an existing harmonic embedding. Experiments on four traffic datasets and a few same-endpoint settings show clear gains over the harmonic embedding alone and results comparable to adding endpoint coordinates.

Critical strengths
  1. The paper identifies a clear weakness of harmonic trajectory embeddings and addresses it with a simple, training-free addition that greatly improves clustering over that baseline on several evaluated datasets.
Critical weaknesses
  1. Incremental technical contribution. The new component appears to be an angular Fourier statistic concatenated with an existing harmonic representations. The idea is straightforward, and the current evaluation does not demonstrate clear advantages (Table 1 shows nearly identical average ARI for ARHR and endpoint-augmented harmonics, while Frechet achieves the best result).

  2. Unclear added value of the angular features. Eq. (4) shows that when a path's angle around the landmark changes in only one direction, these features depend only on its start and end angles and net angular change. On rdb5, adding endpoint angles to the harmonic embedding already matches ARHR's 0.999 ARI. The paper does not compare these simpler features across the main benchmark or report how often and how much paths reverse their angular direction.

Questions for authors
  1. What is the Lyon benchmark? Please describe how the trajectories are generated and how the tasks and route classes are specified.
  2. On the museum data, h+P1 (0.920) clearly beats h+a (0.691). On Lyon, adding h to a lowers ARI from 0.734 to 0.536. Please explain this.
  3. Would ARHR apply to city-scale road networks (e.g., Porto or T-Drive, the kind of data t2vec and TrajCL target), where there is no natural central landmark? If not, the claims should be narrowed to local intersection-scale/landmark-centric clustering.
  4. Table 13 lists ARHR feature extraction at 2.3 s, much longer than some other baselines. Please explain this and clarify the "low per-trajectory feature-extraction cost" claim.
Reviewer_dpgjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies trajectory clustering using topological route representations. Authors proposes Azimuth-Resolved Harmonic Representation (ARHR), which augments cumulative harmonic embeddings with landmark-relative angular activity to better distinguish trajectories with similar harmonic responses but different orientations. Experiments on several real-world traffic datasets report dataset-dependent improvements over the flattened-harmonic baseline, together with weighting and angular-order ablations.

Critical strengths

S1. The paper is generally well motivated and explains why orientation information can complement topology-based trajectory representations.

S2. The method includes analytical characterization and ablation studies, and reports improvements on several real-world traffic datasets.

Critical weaknesses

W1. ARHR mainly augments an existing cumulative harmonic trajectory representation with a compact landmark-relative angular summary, making the contribution closer to a representation enhancement than a fundamentally new clustering framework.

W2. ARHR does not consistently outperform existing baselines in the main absolute-route benchmark (shown in Table1); its primary gains are over the flattened harmonic baseline, while simpler methods such as Fréchet distance achieve comparable or better overall performance.

Questions for authors

Q1. What is the core novelty of ARHR beyond adding angular information to harmonic representations?

Q2. Why is ARHR preferable when simpler baselines such as Fréchet achieve comparable or better results?

Reviewer_2NVFOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARHR, which augments the flattened Hodge-harmonic trajectory embedding of Frantzen et al. with low-order circular Fourier statistics of landmark-relative azimuthal activity, aiming to resolve “orientation ambiguity” where trajectories share similar cumulative harmonic responses. It derives azimuth-shift equivariance and a telescoping reduction of the coefficients, and evaluates route clustering on openDD, Swiss Roundabouts, 3DJunction1 and DLR-UT, plus matched-endpoint settings (controlled, Lyon synthetic, museum UWB). It also reports ablations on activity weights, angular order and harmonic configuration.

Critical strengths

S1. This paper proposes ARHR, a compact trajectory representation that incorporates landmark-relative angular information into harmonic embeddings to address orientation ambiguity and improve route discrimination.

S2. This paper provides theoretical analysis of the proposed representation, establishing its azimuth-shift equivariance and sufficient conditions for distinguishing trajectories with identical harmonic embeddings.

S3. The proposed method is evaluated on four real-world traffic datasets, with extensive comparisons against traditional geometric methods, topological approaches, and learning-based baselines.

Critical weaknesses

W1. Table 2 is the only evidence that angular statistics add information beyond endpoints, yet its settings are small or synthetic (n=24, Lyon synthetic, n=108). The default z₁ = h+a underperforms its own components or controls (Lyon: 0.536 vs a alone 0.734; Museum: 0.691 vs h+P1 0.920). Moreover, no geometric or learned baselines are reported in these settings, although Fréchet-type distances are expected to excel on same-OD routes with different interiors.

W2. In Table 1, Fréchet matches or exceeds ARHR on average (ARI 0.929 vs 0.926; NMI 0.968 vs 0.946), and Yin [h; E_xy] achieves comparable ARI with higher NMI. In Table 13, Yin reaches the same ARI as ARHR on rdb5 (0.999) with about 77× lower feature-extraction time (0.03 s vs 2.3 s). Since the absolute-route labels are largely determined by entry/exit, Table 1 provides no clear reason to prefer ARHR.

W3. Under the default weight wᵢ=|δᵢ|, the coefficients telescope and are fully determined by the endpoint and turning-point azimuths (Eq. 4, App. A.1), so ARHR is essentially a Fourier encoding of these bearings. The authors explicitly decline to compare against an endpoint+turning-point baseline (Sec. 4.2.2), but informational equivalence does not imply equivalent clustering performance. This baseline is needed to establish what the proposed encoding itself contributes.

W4. The landmark is defined by a different rule for each dataset (island centroid, fitted center, vertex-cloud median, junction core median; Table 12), and azimuth statistics depend directly on it. The paper provides neither a sensitivity analysis to landmark displacement nor guidance for choosing a single landmark on multi-hole maps.

Questions for authors

Please see details in Weakness section.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.