acceptodds
Withdrawn from ICLR 2027

Intersection-aware Tangent Decomposition for Manifold Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Clustering intersecting manifolds requires distinguishing nearby observations whose generating branches have different tangent directions. We study when local branch representations can make such clustering possible, and present Intersection- aware Tangent-decomposition Clustering (ITC) as a practical construction mo- tivated by that analysis. The abstract theory identifies sufficient conditions for recovery: a local residual margin controls branch assignment, and a graph perturba- tion bound tolerates arbitrary tangent errors on a vanishing fraction of observations. For balanced orthogonal line segments, we establish a nonempty fixed-radius regime in which imperfect representations and asymptotically optimal two-means rounding yield vanishing misclassification error. This provides an existence re- sult for recoverable intersecting-manifold configurations, together with stability beyond exact tangents. A small-angle oracle example shows why graph recov- erability is a separate requirement. The practical method fits affine branches at spatially distributed anchors, assigns one projector per observation, and constructs a radius-supported affinity. Its finite fitting and spectral routines provide a concrete implementation of the design principles, evaluated empirically; the existence guar- antee concerns the stated abstract realization. A further conditional result covers the degree-rescaled embedding used by the backend.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
multi-manifold clustering, intersecting manifolds, tangent space decomposition, spectral clustering, manifold learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_BoCVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the problem of intersecting manifold clustering. Since different manifolds may contain spatially close observations near their intersections, traditional methods based on Euclidean distances or a single local PCA representation struggle to distinguish samples from different manifold branches. To address this issue, the authors propose Intersection-aware Tangent-decomposition Clustering (ITC). The proposed method first employs local principal component analysis (PCA) to identify potential manifold intersection regions. It then fits multiple affine subspaces around selected anchors and assigns a tangent projector to each observation based on reconstruction residuals. Subsequently, a radius-supported affinity graph is constructed by combining spatial distances and tangent discrepancies, followed by spectral clustering to obtain the final clustering results. Experimental results demonstrate that ITC achieves high clustering accuracy on most of the evaluated datasets.

Critical strengths
  1. Meaningful research problem. The paper addresses the challenging problem of clustering intersecting manifolds, which is meaningful and relevant to manifold clustering research.

  2. Competitive experimental results. The proposed method achieves competitive clustering performance on the evaluated datasets. The authors also provide an analysis of failure cases, which helps clarify the limitations of the method.

Critical weaknesses
  1. Insufficient clarity in problem motivation and methodological novelty. The presentation could be improved. In particular, the paper should more clearly articulate the specific problem it aims to address and distinguish the proposed method from existing approaches. 
  2. Gap between theoretical guarantees and practical implementation. The theoretical results rely on strong assumptions regarding local representation accuracy and optimization quality, which are not guaranteed by the practical algorithm.
Questions for authors
  1. The proposed framework consists of several established techniques, including local PCA, DBSCAN, multiple affine subspace fitting, residual-based assignment, and spectral clustering. Could the authors further clarify the fundamental differences between ITC and existing approaches based on local multi-model fitting, and highlight the key methodological novelty of the proposed framework? 
  2. Could the authors provide additional theoretical or empirical evidence that the finite-restart affine fitting procedure satisfies the representation accuracy assumptions required by the theoretical results?
Reviewer_eoPyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This study tackles the problem of differentiating intersecting clusters. Under a manifold perspective, in a neighborhood near an intersection between two components, Euclidean proximity and a single local covariance are not sufficient to separate them. The paper proposes to tell them apart by estimating the local direction (tangent) of the manifold each point came from. For this, the ITC algorithm uses local PCA to find crossing candidate neighborhoods, groups those with DBSCAN, then fits several linear models around anchor points within each group, assigning to each point the tangent given by its best fitting model. The resulting tangents are used to build a similarity graph where points close in space are connected if they have similar tangents. The end result is produced by running spectral clustering on that graph. The paper offers a theoretical guarantee for the special case of two orthogonal lines crossing at a point, and shows that ITC outperforms previous methods on seven synthetic datasets.

Critical strengths

The paper tackles a notoriously difficult problem, achieving better results on synthetic benchmarks than previous approaches.

Critical weaknesses
  1. The theory is not strong enough to support the method as a practical approach. The statistical guarantees apply to idealized scenarios (noiseless, orthogonal segments, fixed radius, etc.) far from what real-world data looks like. Although the paper claims that the theory motivates the design, it does not cover the finite fitting, selection of K, or spectral steps actually used by the algorithm in the experiments. Thus, one is left with the empirical findings to judge whether the approach makes sense. However, all of the manifolds used in the benchmark have low intrinsic dimension (1 or 2, with 0 or 1-dim. intersections). Ambient dimension is at most 3, with no isotropic ambient noise. All the manifolds have low curvature. The theorem covers only straight lines with orthogonal crossing; the algorithm is written for general dd but is only shown to work empirically up to d=2d=2. Thus, it is hard to assess whether the assumptions used in the paper would hold in higher-dimensional or more realistic settings. In addition, the pipeline concatenates multiple standard components (local PCA, DBSCAN, affine fitting, spectral clustering) but the theory was not extended to the full procedure.

  2. Clarity and reproducibility: already in the abstract, questions arise: what are “balanced orthogonal line segments”? What does “Its finite fitting and spectral routines...” refer to? The heavy use of self-defined jargon for naming simple steps in the algorithm makes it harder to follow (e.g., “candidate mixed neighborhoods”, “medoid–farthest-point anchors with covering tolerance”, “single-node position–tangent graph”, “A1 spatial scale”, “spectral backend and theoretical realization” could use a plain description of the operation instead). Several steps in the Method are not sufficiently explained or justified; for example, the reader must infer what R_d(i) represents; ρR\rho_R and γ\gamma are used in section 3.1 but never defined in the main text nor justified in the appendix. σT\sigma_T from section 4 is fixed at 0.2 without justification either. Moreover, the approach relies strongly on the output from local PCA and DBSCAN to decide which neighborhoods are intersections and which points share an anchor, bringing with them their own assumptions and hyperparameters. But that is never mentioned in the abstract or introduction, and the sensitivity to their hyperparameters (τR\tau_R, rr, qmin⁡q_{\min}) is not discussed or analyzed. Finally, the paper provides no source code, which would be important for clarifying details of these steps and ensuring reproducibility.

Questions for authors
  1. Fig. 2g is labeled “Swiss Roll + Plane” (as in App C.1), but the scatter plot shows a swiss roll intersected by an annulus, which is topologically distinct. Which is correct?

  2. Could the source code be provided, as well as scripts to reproduce the reported results?

  3. “Runs without valid clustering outputs are recorded as failures” -- I found no analysis of what causes the execution failures seen on the 2-d manifolds. In a real-world scenario where ground truth is available, is it possible to distinguish a failure from a legitimate run that simply returned the wrong partition? In what cases might a component have no valid anchors?

  4. How to select the hyperparameters qmin⁡q_{\min}, τR\tau_R and rr? The paper offers no discussion of how sensitive the method is to these critical neighborhood parameters, and using a fixed radius is widely known to be prone to failure in real-world, non-uniformly sampled data.

  5. One body of work that is arguably missing from the related-work section is multi-structure robust fitting — J-linkage (Toldo & Fusiello, ECCV, 2008), T-linkage (Magri & Fusiello, CVPR, 2014), and RansaCov (Magri & Fusiello, CVPR, 2016) all fit multiple local models and reassign points by residual, which is the same mechanism as ITC’s anchor fitting step, just applied in a different domain (geometric model fitting rather than manifold clustering). Could the paper situate itself against this line of work?

Reviewer_Nz8XOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper attempts to address the problem of clustering data lying on multiple manifolds, and in particular the case when the manifolds intersect. An algorithm is proposed, together with some theoretical investigations. Synthetic experiments are conducted to demonstrate the effectiveness of the proposed method.

Critical strengths

N/A

Critical weaknesses

Clarity

  1. The writing of the paper has major problems that hinders readability. Below are a few examples:
    1. The word "branch" is used throughout the paper, such as "generating branches", local branch representations", "branch fitting", "branch counting", "noisy branch". However, it is not defined in the beginning of the paper, if it is defined at all. By "branch", do you mean manifold? or a branched manifold?
    2. Abstract: "Clustering intersecting manifolds" do you mean clustering data samples from intersecting manifolds?
    3. The preamble of section 3 started by defining notations. While the text laid out the symbols, it was not mentioned what the symbols mean. For instance, what is \xi_i -- is it noise? What is y_i and M_{y_i}?

Significance

  1. Clustering data from manifolds is a classic problem. When the manifolds are linear, methods exist to deal with the case of intersecting manifolds, such as algebraic subspace clustering, and sparse subspace clustering (Elhamifar & Vidal, 2013). When the manifold is not linear, the latter approach has been extended (Elhamifar & Vidal, 2011). It has been shown that such an approach works well on large-scale datasets [1]. Why are we not satisfied with the latter approach, both conceptually and emprically?
  2. To demonstrate the impact of the proposed method, one needs to consider large-scale realistic datasets. For instance, one may run the proposed method on CLIP (or other vision language model) embeddings of images from ImageNet, and see if different clusters intersect. However, the experiments in this paper focuses on synthetic data.

[1] Geometric Analysis of Nonlinear Manifold Clustering, NeurIPS 2024.

Questions for authors

Please see the questions above.

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

Related papers

Discussion (0)

Sign in to comment.