acceptodds
Under review as a conference paper at ICLR 2027

Multiscale Euclidean Network Trajectories: Second-Moment Geometry, Attribution, and Change Points

Abstract

Understanding a dynamic network requires more than locating when it changes: we also want to know the magnitude and structural direction of change and which nodes contribute. Unfolded spectral embedding places network snapshots in a joint representation, but the underlying rectangular latent factorization is invariant under general linear transformations. Such transformations preserve edge probabilities while potentially distorting Euclidean second moments. We introduce Multiscale Euclidean Network Trajectories (MENT), whose central estimands are pairwise second-moment operators of the dynamic latent positions. Under an isotropic normalization of the shared anchor latent positions, these operators are identifiable up to simultaneous orthogonal conjugation. They yield a global trace variation distance and, through an aggregated operator, shared orthogonal directions for mode-wise distances. The squared global distance decomposes exactly across these modes. Classical multidimensional scaling turns the distances into global and mode-wise trajectories of time. A modified unfolded spectral estimator consistently recovers the operators and induced trajectories, and we establish bounds connecting node attributions to trajectory displacements and trajectory estimation to change point localization. Synthetic experiments validate geometric recovery and show improved detection of mode-specific changes against baselines. Analyses of two real networks illustrate interpretable mode- and node-level descriptions of temporal evolution.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
dynamic networks, spectral embedding, identifiability, latent position models, change-point detection

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_Di2dOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Multiscale Euclidean Network Trajectories (MENT), a framework that maps the snapshot-level evolution of a dynamic network into comparable geometric trajectories over time to simultaneously address three core questions: when the network changes, along which latent structural directions it changes, and which nodes contribute most to the changes. The core theoretical innovation of this work lies in resolving the coordinate-stretching ambiguity under general linear transformations GL(d)GL(d) inherent to the rectangular latent factorization of existing unfolded adjacency spectral embedding (UASE). By imposing an isotropic normalization on the anchor latent positions, MENT reduces the ambiguity of the pairwise second-moment operators to a shared orthogonal transformation, thereby ensuring the statistical identifiability of both temporal Euclidean distances and multi-mode orthogonal decompositions.

Critical strengths

1.By extracting a shared orthonormal eigenbasis from an aggregated second-moment operator, the squared global trace variation distance decomposes exactly into the sum of squared mode-wise distances, cleanly reconciling global evolutionary summaries with fine-grained structural disentanglement.

2.The real-world case studies maintain commendable academic restraint, framing the empirical findings strictly as descriptive network evolution patterns without overinterpreting high node attributions as personal responsibility or causal mechanisms.

Critical weaknesses
  1. The paper presents Theorem 3.4 as a change-point localization result, but the theorem concerns a single-knot least-squares estimator based on a piecewise parametric trajectory. The actual multiple-change-point procedure used in the main Dataset 2 experiment is different. It fits local linear trend models to the estimated mode-wise trajectories, extracts level and slope innovation scores, and then fuses candidates across modes and change orders. The paper explicitly acknowledges that the multiple-change-point procedure is evaluated empirically rather than covered by Theorem 3.4. This is not a correctness issue, but it creates a gap between the theoretical localization result and one of the strongest downstream empirical claims of the paper. I think the paper should more carefully distinguish between: consistency of the estimated trajectories and theoretical guarantees for the specific multiple-change-point detector.

  2. My main concern is the empirical generality of the result. Both synthetic datasets are constructed with three orthogonal latent modes, while a central feature of MENT is precisely to recover persistent orthogonal directions of temporal variation. Dataset 2 then places multiple change points across these different modes, which is exactly the setting in which a mode-wise method should be expected to have an advantage over global summaries.

The resulting performance is impressive: at K=6, MENT reaches F1 = 0.965 with timing MAE =0.10, substantially outperforming the competing methods. However, the current experiment primarily establishes that MENT works very well when temporal changes occur along a fixed orthogonal latent decomposition closely matching the proposed representation.

I would be substantially more convinced by an additional setting in which this alignment is deliberately violated. The two real-world examples are useful for interpretability, but they do not provide the same type of quantitative ground-truth comparison against competing methods.

Questions for authors
  1. How dependent is the empirical advantage of MENT on the existence of a fixed orthogonal latent modal structure? Could the authors provide an experiment with non-orthogonal, rotating, or otherwise misspecified latent directions to test robustness beyond the synthetic setting that closely matches the proposed decomposition?

  2. Can the authors clarify the theoretical status of the multiple-change-point procedure used in Dataset 2? Theorem 3.4 analyzes a single-knot least-squares estimator, whereas the reported results use innovation scores and candidate fusion. Is there a stability or localization result that can be established for the actual procedure?

  3. What should be regarded as the central methodological novelty after canonicalization? My reading is that the key new ingredient is the identifiable canonical second-moment geometry and its statistical recovery, whereas the orthogonal decomposition, spectral ordering of modes, and CMDS steps are largely standard once that object is available. I would appreciate a clearer positioning relative to whitening/canonicalization, PCA-style shared modes, and Euclidean Mirror/UASE.

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

The paper proposes a variation on unfolded spectral embedding of a dynamic network which identifies positions up to orthogonal transformation, and builds global, nodewise, and directionwise distances, trajectories with changepoint detection.

Critical strengths

Exposing and analysing this canonical representation of the positions is an interesting idea. The consequences for decomposing trajectories into identifiable coordinate-wise contributions is potentially useful.

Critical weaknesses

Is the procedure not the same as that proposed for Intensity Profile Projection (IPP) [1]? (up to scaling by n^{-1/2}, and assuming a histogram kernel estimator). This paper is not cited, which is a significant omission if so.

More generally, the level of referencing is not acceptable. The authors don't even provide a reference for the random dot product graph.

Some of the theorems seem very elementary, e.g. is there more to 2.4 than Pythagoras?

Because of the above, I am not sure if there is a substantial novelty in this paper.

[1] A. Modell, I. Gallagher, E. Ceccherini, N. Whiteley and P. Rubin-Delanchy. Intensity Profile Projection: A Framework for Continuous-Time Representation Learning for Dynamic Networks. Advances in Neural Information Processing Systems 36 (NeurIPS 2023). arXiv:2306.06155.

Questions for authors

Does the procedure differ from IPP above?

If not, could you clarify what remains novel and unexpected within the present paper, once we understand (after IPP) that the embeddings are just standard PCA-style projections of the connection profiles?

Reviewer_jqUZOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MENT, a framework for learning temporal representations of dynamic graphs from unfolded spectral embeddings while addressing the (GL(d)) non-identifiability of the underlying rectangular latent factorization. The main idea is to impose an isotropic normalization on a shared, time-invariant anchor latent factor, reducing the ambiguity from general linear transformations to a single shared orthogonal transformation. The resulting pairwise second-moment operators are then used to define global and mode-wise temporal distances, which are mapped through classical MDS into trajectories of network evolution and subsequently used for tasks such as change-point detection. Overall I find the paper very interesting and well written tackling this very exciting, in my opinion, problem.

Critical strengths

The paper targets a real, well-known, and important issue concerning the interpretation of Euclidean geometry obtained from unfolded spectral factors, while providing a principled operator-to-distance-to-trajectory construction. The problem is particularly relevant for temporal graph representation learning, where obtaining representations that permit robust and identifiable extraction of evolving structure remains challenging. I believe this framework may also motivate follow-up work on interpretable temporal graph embeddings and potentially on dynamic community structure. Related approaches based on tensor decompositions, such as PARAFAC/CP, have been used extensively for extracting temporal network structure, but impose particular low-rank multilinear assumptions that can limit the types of dynamics they represent. In this respect, the combination of unfolded spectral representations, an explicit treatment of identifiability, and accompanying theoretical analysis provides an interesting alternative. I found the paper well motivated and enjoyable to read.

Critical weaknesses

There has been another family of methods that address such non-identifiability issues, mostly for static graphs, via the use of simplex/NMF/mixed-membership properties. I therefore would like the authors to better position the isotropic canonicalization relative to such literature. For example, HM-LDM (Nakis et al.) constrains latent positions to a simplex to address latent-space non-identifiability, while MMSBM/SNMF work (e.g., Mao, Sarkar & Chakrabarti; Huang & Fu) establishes identifiability under pure-node/separability or weaker geometric conditions, often up to permutation. Could an analogous simplex constraint on the anchor (X) reduce the present (GL(d)) ambiguity directly to permutations? If so, what is gained specifically by the isotropic (GL(d) -> O(d)) construction relative to these alternatives? A discussion of the assumptions/expressiveness trade-off would substantially clarify the novelty. I find it particularly interesting if the direction that MENT follows leads to more expressive embeddings which is what I would expect.


MENT does not identify the latent representation itself uniquely but it remains identifiable only up to a shared (O(d)) transformation. This is sufficient for the proposed invariant second-moment geometry, but what does it imply for interpretation of individual latent directions and signed node-level attributions? think the paper should distinguish more clearly between identifiability of the proposed geometric quantities and identifiability of the underlying latent representation.


The theory assumes a common anchor structure with anchor isotropy. It would be useful to understand how this choice constrains the expressiveness of the provided model against a completely free version. This could also be made more clear with additional downstream tasks, such as link prediction or network reconstruction.

In terms of the empirical gains as the proposed pipeline contains multiple components, it would be good to show more specifically how much of the observed gains come from the proposed geometry, probably through some ablations.

Finally, there is no complexity analysis of the method which of course is at least quadratic due to fact of the use of MDS. There have been many methods for scaling such operations so it is not necessary to add experiments regarding this, but a paragraph discussing methods of scaling the approach are necessary.

I will be comfortable to move towards a clear acceptance if the authors address successfully my concerns.

Questions for authors
  1. Could the GL(d)GL(d) ambiguity instead be resolved via simplex/nonnegative constraints on XX, as in MMSBM, identifiable NMF/SNMF, or HM-LDM, and how would that compare with the proposed isotropic construction?

  2. How restrictive is the shared-anchor assumption �2��2�, and what types of temporal evolution, independently evolving memberships or latent coordinates, cannot be represented?

  3. Is anchor isotropy primarily a substantive modeling assumption or a canonical whitening/gauge choice, and why is it preferable to simplex, nonnegativity, or other identifiable factorizations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.