{"id":"776605f8-5f24-4a0e-a461-b6b7101f2967","arxiv_id":"2502.02185","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GenKSC couples kernel spectral clustering with a generative decoder and fixed simplex cluster codes, enabling traversal-based visualization of cluster-defining features.","lead":"The paper introduces GenKSC, a clustering model that combines kernel spectral clustering with a generative decoder, allowing users to generate images that exaggerate each cluster's defining features. Tested on MNIST and FashionMNIST, traversals along cluster directions produce recognizable, cluster-specific image variations, aiming to make deep clustering interpretable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The decoder is trained only on projected training points; extrapolating U e* along a cluster direction has no supervision, so the claim that traversals exaggerate cluster features rests on an untested assumption of out-of-distribution generalization by psi.","rationale":"The reader's weakest_assumption correctly identifies the inverse mapping's generalization as the key risk. I agree, with a refinement: the issue is not only generalization error but the absence of any training signal that would make the decoder's extrapolation semantically meaningful. The model has no latent prior (unlike a VAE) and no adversarial realism constraint (unlike ClusterGAN), so the decoder is free to produce arbitrary outputs outside the training manifold. The paper's sole evidence is a small set of images (Figures 2-3) with no quantitative validation that the generated points belong to the target cluster or that the traversals correspond to cluster-discriminative features. A concrete test with an independent classifier would settle whether the central claim lands. If the test fails, the novelty argument collapses; if it passes, the concern is resolved. The reader's CONDITIONAL verdict remains appropriate, possibly with the addition of this required experiment as a condition.","tokens_in":4472,"tokens_out":6951,"duration_ms":72697,"concrete_test":"Train GenKSC on MNIST012 with the paper's settings (s=10, k=3, eta_rec=eta_cl=1, cluster loss after 10 epochs). Let e_c be the mean score vector of training points assigned to cluster c. Generate x*(t)=psi(U (t e_c)) for t=0, 0.5, 1, 1.5, 2, 3. For each t, feed all generated images (across the three clusters) to a pre-trained MNIST classifier and record the fraction with correct-class softmax probability >0.9. Also generate images along random directions U r with ||r||=||e_c|| as a control. If the fraction drops below 50% for t>1, or random directions produce visually similar 'exaggeration' in a blind comparison, the extrapolation claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that extrapolating in latent space yields points that 'emphasize or exaggerate distinctive cluster features' (Conclusion). For this to hold, the inverse map psi (Section 3.2) must produce realistic, cluster-specific images when evaluated at U e* for e* far beyond the training projections (Section 3.3). But the reconstruction loss Lrec only supervises psi on the projected training features U U^T phi(x_i); there is no prior, Lipschitz penalty, or adversarial loss constraining psi outside that set. Thus the decoded traversals in Figures 2-3 could be arbitrary decoder artifacts. The paper provides only qualitative images, with no quantitative check that generated points are classified as the target cluster or that the direction is aligned with discriminative features. This is the load-bearing assumption because the novelty claim rests entirely on the meaningfulness of extrapolated traversals.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Generative Kernel Spectral Clustering (GenKSC), a model that combines kernel spectral clustering with a generative autoencoder-like architecture. The method trains a parametric feature map and an approximate inverse map by minimizing a weighted variance objective together with a reconstruction loss and a cosine-distance cluster loss whose target directions are fixed as vertices of a regular simplex. After training, new points are generated by decoding extrapolated score-vector locations, producing traversals intended to exaggerate cluster-specific features. Experiments on a three-digit subset of MNIST and on FashionMNIST present qualitative latent-space traversals, and the authors claim that these traversals reveal and exaggerate distinctive cluster characteristics.","tokens_in":4685,"tokens_out":3258,"duration_ms":36821,"significance":"If the central claim holds, GenKSC would make a useful contribution by coupling spectral-clustering-style embeddings with generative decoding, yielding explorable latent spaces that are both clustering-oriented and interpretable. A notable strength is the explicit combination of a weighted-variance KSC objective with learned deep feature maps, an idea with clear potential for interpretable deep clustering. However, the current evidence is largely qualitative: the manuscript contains no quantitative clustering metrics, no baseline comparisons, no error bars, and no validation that extrapolated traversals remain faithful to the target cluster. The significance of the method therefore remains plausible but unestablished at the level expected for a journal publication.","major_comments":[{"comment":"The objective includes the term ||phi(x_i)||_2^2 described only as \"regularization on feature representations.\" This term is not a constant: D_ii itself depends on phi through K, so the term modifies the weighted variance problem in a nontrivial way, and no justification or practical role for it is given. The authors should either derive this term from a principled constraint or remove it, and they should explain why no weighting hyperparameter is attached to it.","section":"Section 3.3, Eq. (3)"},{"comment":"The equivalence between the eigendecomposition in Eq. (1) and the weighted PCA formulation in Eq. (2) is asserted in a single sentence referring to stationarity conditions, but no derivation is provided. Because the paper then uses a learned feature map and adds reconstruction and cluster losses, it is important to state precisely under which conditions the equivalence holds and how the centering assumption in footnote 1 affects it. A short derivation or a precise reference to the corresponding primal-dual result is needed.","section":"Section 3.1, after Eq. (2)"},{"comment":"The cluster directions are defined a priori as vertices of a regular simplex, and the cluster loss minimizes the cosine distance to these fixed codes. Consequently the \"traversals along cluster directions\" in Figs. 2 and 3 follow predefined directions rather than directions discovered from the data. The interpretability claim is therefore partly circular: the model is forced to align with the fixed codes, and then those same codes are used to interpret the traversals. To support the claim, the authors should demonstrate that the resulting cluster assignments are meaningful, for example by reporting clustering accuracy or NMI with respect to known labels and by comparing with a standard KSC baseline.","section":"Section 3.2, cluster loss"},{"comment":"The inverse map psi is trained only through L_rec on the projected training points U U^T phi(x_i). The generation step x* = psi(U e*) evaluates psi at points that can lie far beyond the convex hull of the training projections, and no mechanism (e.g., Lipschitz regularization, adversarial training, or density modeling) constrains psi in that region. The claim that extrapolated traversals \"exaggerate distinctive cluster features\" therefore rests on an unvalidated out-of-distribution generalization assumption. The authors should provide quantitative evidence that extrapolated images are classified as the intended cluster, or otherwise measure how cluster fidelity degrades as the traversal moves outward.","section":"Section 3.3 and Section 4.2"},{"comment":"The experiments report no quantitative clustering metrics, no error bars, and no baseline comparisons. For a clustering paper, this is a serious omission: the reader cannot assess whether GenKSC produces well-defined clusters in any measurable sense. In addition, Section 4.1 states that FashionMNIST is trained with k=10, while Fig. 3 reports traversals for \"6 clusters\" and Section 4.1 also mentions hyperparameter tuning based on an \"average membership strength criterion\" that is not defined. These inconsistencies and omissions should be addressed with a full experimental evaluation.","section":"Section 4, Experiments"}],"minor_comments":[{"comment":"There is a typographical line-break artifact in the abstract: \"interpretabili ty\" should be \"interpretability.\" Please proofread the text.","section":"Abstract and Section 1"},{"comment":"The relation e_i = D_ii h_i Lambda is stated without defining the dimensions of the score vectors and embeddings clearly. In particular, the later notation e* in Section 3.3 and the operation U e* should be explained more carefully, since U maps from feature space to score space while U e* maps from score space back to feature space; a consistent convention would help the reader.","section":"Section 3.1, notation"},{"comment":"The cluster loss is applied only to the first k-1 components of the score vectors, while the model allows s > k-1. The paper should discuss why the remaining components are excluded and what role they play in the reconstruction and in the generation of intra-cluster variations.","section":"Section 3.2, cluster loss"},{"comment":"The hyperparameter selection for FashionMNIST relies on an \"average membership strength criterion\" referenced as in [8], but no definition is given in the text. A one-sentence explanation would make the experimental setup self-contained.","section":"Section 4.1, hyperparameters"},{"comment":"Footnote 2 points to reference [9] for a detailed comparison of primal and dual formulations, but the citation appears after the formula and the connection to the present paper is not elaborated. Please clarify the intended reference and its relevance.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This manuscript reads as a workshop contribution (ESANN 2025) rather than as a full journal paper. The central idea is appealing, but the missing derivations, the unexplained regularizer, and especially the absence of any quantitative validation of the extrapolated traversals are substantial gaps. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also encourage the authors to make their code and trained models available, since the qualitative figures alone cannot support the novelty claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read of Winant, Achten, and Suykens, \"Generative Kernel Spectral Clustering.\" The short version: it's a legitimate incremental contribution — KSC with a learned feature map, a reconstruction loss, and fixed simplex cluster codes, trained jointly so you can decode along cluster directions. On MNIST012 and FashionMNIST the traversals look reasonable. But the paper's one big claim, that extrapolating in latent space \"emphasizes or exaggerates\" cluster-defining features, is supported only by eyeballing images, and the mechanism that would make it true is never examined.\n\nThe synthesis is genuinely new relative to what they cite: KSC gives the line structure, Stiefel-RKM gives the encoder-decoder setup, ClusterGAN gives the clustering-in-latent-space idea, and nobody combines them with decoding along cluster directions. The writing is honest about the lineage. The math is mostly standard — the \"equivalence\" between problems (1) and (2) is stated laconically, but it's the standard weighted-kernel-PCA stationarity argument from Alzate and Suykens, so I'd call that a presentation gap, not an error.\n\nThe real problems, in order. First, the stress-test note is right: ψ is trained only on projected training features U U^T φ(x_i). When you decode U e* for e* beyond the training range, nothing supervises the decoder. No Lipschitz penalty, no prior, no adversarial loss. So the \"exaggeration\" result could be decoder artifact, and the paper gives no quantitative check that the generated points actually belong to the target cluster or that the direction aligns with discriminative features. Second, there is no quantitative evaluation at all — no clustering accuracy, NMI, ARI, no baselines (KSC, ClusterGAN, plain autoencoder), no error bars. For a paper whose claims are \"well-defined clusters\" and \"interpretable representations,\" the evidence is images and scatter plots only. Third, the ||φ||² term in Eq. (3) appears with no explanation; it's doing real work (the weighted variance term is scale-invariant in φ, so the regularizer picks a scale) but the paper doesn't say so. That's fixable in a sentence.\n\nThe circularity concern about fixed simplex codes is present but mild: the partition still comes from the data; the codes just kill the rotation ambiguity. I wouldn't hang a rejection on that.\n\nWho's this for? People working on interpretable clustering or the RKM/KSC line. It's an honest, buildable-on idea with thin evidence. It deserves a serious referee, and my read is a conditional: require quantitative evaluation, one baseline, and either a regularity argument for the decoder or a tempered claim about extrapolation.","headline":"A plausible incremental extension of KSC with a generative decoder, but the claim that extrapolated traversals exaggerate cluster features rests on unexamined out-of-distribution decoding and on image-only evidence.","tokens_in":5196,"tokens_out":6252,"would_cite":true,"duration_ms":58965,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Clustering that can generate and exaggerate its own cluster images.","keywords":["kernel spectral clustering","generative modeling","interpretable clustering","latent space traversal","deep clustering","reconstruction loss","explainable AI"],"falsifier":"Take a dataset with known labels, generate decoded images at increasing distances along a cluster direction, and classify them with a pretrained classifier; the central claim fails if class identity collapses or shifts before characteristic exaggeration appears.","tokens_in":4296,"feed_emoji":"🖼️","tokens_out":5861,"duration_ms":54861,"temperature":0.7,"pith_summary":"The paper proposes GenKSC, a clustering model that combines kernel spectral clustering with a learned encoder-decoder so that clusters are not only assigned but also visualized. The model trains a feature map and an inverse map together, adding reconstruction and cluster-direction losses to the spectral variance objective. A user can pick a point in the score space, decode it, and obtain an image that represents or exaggerates a cluster's distinguishing features. The authors demonstrate on MNIST and FashionMNIST subsets that traversals along cluster directions reveal what the clustering is based on. The point of the work is to close the gap between deep clustering accuracy and interpretability.","feed_headline":"New clustering model draws images that exaggerate each cluster","feed_subtitle":"GenKSC turns spectral clustering into a generative latent space you can walk through and inspect.","key_machinery":"The central object is the score-vector latent space $e_i = U^\\top \\varphi(x_i; \\theta_\\varphi)$ together with the inverse map $\\psi(U e; \\theta_\\psi)$. The training objective combines a negative degree-weighted variance term that creates the spectral clustering structure, a feature-map regularization, an autoencoding reconstruction loss, and a cosine-distance cluster loss that aligns the first $k-1$ score components toward $k$ simplex vertices. This combination is what makes the latent directions stable enough that decoding far extrapolations along a cluster direction stays cluster-specific.","core_discovery":"GenKSC shows that the line-structured eigenspace of kernel spectral clustering can be turned into a generative, explorable latent space. The feature map $\\varphi$ and approximate inverse $\\psi$ are learned along with the projection $U$, and the objective augments weighted variance maximization with a reconstruction term and a cosine cluster loss that pushes score vectors onto predefined directions, namely the vertices of a regular simplex. Because $U$ is orthonormal, one can select any score $e^*$ and decode it as $x^* = \\psi(U e^*)$; following a cluster direction beyond the data range generates images that exaggerate the cluster's characteristic features, while higher latent components expose intra-cluster variation.","pith_inferences":["Traversal distance could serve as a cluster-quality diagnostic: if extreme extrapolations lose class identity or bleed into another cluster, the clusters are not cleanly separated.","The generation step could be turned into a counterfactual tool, interpolating between cluster codes to show which features must change to move an input from one cluster to another.","Quantifying generation with classifier agreement or distribution distances would test whether visual exaggeration is actually class-preserving, since the paper's evidence is visual only.","The method might extend to non-image tabular data, where cluster directions would be less visual but could still be decoded as synthetic examples; this extension is untested."],"forward_implications":["Users can generate a representative image for a cluster by decoding a point on that cluster's direction, giving a direct visual definition of each cluster.","Extrapolating beyond the farthest training point along a cluster direction exaggerates the cluster's defining traits, so cluster boundaries and failure modes become visible.","Intra-cluster variation can be explored by moving along higher latent dimensions beyond the first $k-1$ directions.","The same loss combination could be reused in semi-supervised or supervised settings, where cluster labels from another model are used to build an interpretable latent space."],"supporting_citations":[{"why":"Supplies the kernel spectral clustering formulation, the weighted PCA eigenspace, that GenKSC builds on.","marker":"[7]"},{"why":"Provides the inverse-mapping generative architecture used to decode latent points.","marker":"[4]"},{"why":"Supplies the adversarial clustering baseline and the convolutional architecture adapted for FashionMNIST.","marker":"[6]"},{"why":"Defines the cosine-similarity cluster quality criterion used for hyperparameter selection.","marker":"[8]"},{"why":"Provides the Cayley-transform optimizer used to enforce the orthonormal constraint on U.","marker":"[10]"},{"why":"Introduces the restricted kernel machine framework from which the feature and inverse-map design is drawn.","marker":"[3]"},{"why":"Gives the primal-dual equivalence that connects the optimization to the KSC formulation.","marker":"[9]"}],"fun_headline_variants":["Generative spectral clustering lets you visualize cluster traits","Walk through cluster space: GenKSC generates exaggerated features","GenKSC: Turn clusters into images by traversing latent directions","Exaggerate cluster features with a generative kernel spectral model","Generate cluster archetypes with GenKSC's explorable latent space"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's decoder, trained on ordinary training images, has to keep producing meaningful cluster-specific images when asked to decode points far beyond the training range along a cluster direction.","fun_headline_variants_meta":{"raw":{"variants":["Generative spectral clustering lets you visualize cluster traits","Walk through cluster space: GenKSC generates exaggerated features","GenKSC: Turn clusters into images by traversing latent directions","Exaggerate cluster features with a generative kernel spectral model","Generate cluster archetypes with GenKSC's explorable latent space"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2721,"prompt_tokens":746,"completion_tokens":1975,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":362,"completion_tokens_details":{"reasoning_tokens":1892}},"tokens_in":362,"tokens_out":1975,"duration_ms":13195,"temperature":1.0,"reasoning_tokens":1892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:01:59.423837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with known labels, generate decoded images at increasing distances along a cluster direction, and classify them with a pretrained classifier; the central claim fails if class identity collapses or shifts before characteristic exaggeration appears.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the kernel spectral clustering formulation, the weighted PCA eigenspace, that GenKSC builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the inverse-mapping generative architecture used to decode latent points."},{"cited_title":"ClusterGAN: Latent Space Clustering in Generative Adversarial Network s","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial clustering baseline and the convolutional architecture adapted for FashionMNIST."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the cosine-similarity cluster quality criterion used for hyperparameter selection."},{"cited_title":"Efficient Riemann ian Optimization on the Stiefel Manifold via the Cayley Transform","cited_arxiv_id":null,"evidence_quote":"Provides the Cayley-transform optimizer used to enforce the orthonormal constraint on U."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the restricted kernel machine framework from which the feature and inverse-map design is drawn."},{"cited_title":"Duality in Multi-View Restricted Kernel Machines","cited_arxiv_id":"2305.17251","evidence_quote":"Gives the primal-dual equivalence that connects the optimization to the KSC formulation."}],"review_version":1}