{"id":"1854a281-9b87-4cf1-883e-0899d4761169","arxiv_id":"2501.01595","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"AHSGC combines an adaptive filter graph encoder with pseudo-label-driven graph edge updates to improve unsupervised clustering accuracy on Salinas, Pavia University, and Trento hyperspectral images.","lead":"This paper proposes AHSGC, an unsupervised graph clustering method for hyperspectral images that builds a superpixel graph, encodes it with a learnable high/low-pass filter, and updates the graph using clustering pseudo-labels. On three standard HSI benchmarks it reports higher overall accuracy than nine baselines, though some reported metric claims conflict with its own tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AHSGC is not best on NMI/Purity in its own Tables IV–V; Section IV-C.1 misreports NMI, undermining the central 'best clustering performance' claim.","rationale":"In good faith, the paper attempts to demonstrate that AHSGC is a state-of-the-art unsupervised HSI clustering method. The central claim is the reported superiority on three datasets across five metrics. For that claim to hold, all five metrics must be correctly computed and the method must actually be best on them. The manuscript's own tables show this is not the case: AHSGC is not the best on NMI on Salinas or PU, and not the best on Purity on PU. The text in Section IV-C.1 also contains an internal inconsistency, listing four numerical values for five named metrics. This is a concrete, verifiable failure of the headline claim, independent of any speculation about the graph-update mechanism. The reader's weakest assumption focused on the pseudo-label self-reinforcement loop in the graph sparsification; while that is a legitimate methodological concern, the factual contradiction in the reported metrics is more directly determinative of the paper's central claim. I therefore select the metric misreport as the single most load-bearing concern. The proposed check is a straightforward re-computation and re-reporting of the ten-run results; if the numbers reproduce as in Tables IV and V, the 'best on all five metrics' statement cannot be sustained. The paper could still be correct on OA, and the method may still be valuable, so a conditional acceptance requiring corrected reporting and honest metric-wise claims is appropriate. This does not move the reader's verdict of CONDITIONAL, since that verdict already anticipated red flags; however, the specific weakness I identify is different from the reader's chosen weakest assumption, hence 'partial' agreement.","tokens_in":20889,"tokens_out":7282,"duration_ms":65554,"concrete_test":"Cross-check the reported NMI and Purity columns in Tables IV and V against the text in Section IV-C.1: if AHSGC NMI (0.8587 on Salinas; 0.5251 on PU) is below EGAE (0.8769; 0.5372) or K-means (0.5529 on PU), and Purity on PU is below EGAE (0.7126), then the central claim of 'best clustering performance' across all five metrics is contradicted. To settle whether this is a typo or a systematic reporting issue, re-run all methods ten times with the stated hyperparameters and report mean ± std for every metric; if the confidence intervals reconfirm AHSGC NMI < EGAE NMI on Salinas and PU, the claim must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim, stated in Section IV-C.1, is that AHSGC achieves the best clustering performance on Salinas across OA, Kappa, NMI, ARI, and Purity. This is contradicted by the paper's own Table IV. The text lists four numbers for five metrics ('83.60%, 81.62%, 77.42%, and 83.68%'), omitting NMI. The table reports AHSGC NMI = 0.8587, while EGAE achieves NMI = 0.8769. AHSGC is therefore not the best on NMI on Salinas. The claimed 0.73% improvement over the 'second-best' uses NCSC (0.8514) rather than the actual second-best EGAE (0.8769), turning the alleged gain into a 1.82 percentage-point deficit. On PU, Table V shows AHSGC NMI = 0.5251, below K-means (0.5529) and EGAE (0.5372), and Purity = 0.6980, below EGAE (0.7126). Thus, even before considering variance or test-set hyperparameter selection, the empirical claim of best clustering performance across all five metrics fails on two of the three datasets. This is an internal inconsistency in the reported results, not merely a disagreement with external benchmarks. As the headline result is the paper's main contribution, this misreport is load-bearing: the claim 'best clustering performance' is false as written, and the corrected conclusion would need to specify which metrics and datasets actually support it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes AHSGC, an unsupervised clustering method for hyperspectral images that combines superpixel-based graph construction, an adaptive-filter graph encoder for low- and high-frequency feature extraction, a self-training clustering decoder based on KL divergence, and a homophily-enhanced structure-learning module that sparsifies and rewires the graph during training. The method is evaluated on Salinas, Pavia University, and Trento, compared against nine clustering baselines, and reported to achieve higher OA, Kappa, NMI, ARI, and Purity on most settings, along with lower computational cost. Ablation studies and hyperparameter sensitivity analyses are also provided.","tokens_in":21297,"tokens_out":4666,"duration_ms":46937,"significance":"If the empirical claims were fully supported, AHSGC would be a practically useful contribution to unsupervised HSI clustering: it combines several plausible ingredients (superpixel graphs, adaptive filtering, dynamic graph sparsification, joint self-training) and the reported runtime/complexity advantages are attractive. The paper also ships an extensive comparison, ablation experiments, and a promise of code release. However, the central 'best clustering performance' claim is internally contradicted by the paper's own tables on two of the three datasets, and the self-training graph-update loop has a circularity that is not addressed. These issues are load-bearing because the headline contribution is precisely the claimed state-of-the-art accuracy and the adaptive graph-updating mechanism.","major_comments":[{"comment":"The text states that AHSGC 'achieves the best clustering performance with 83.60%, 81.62%, 77.42%, and 83.68% in terms of OA, Kappa, NMI, ARI, and Purity' on Salinas and claims a 0.73% improvement on NMI. This is contradicted by Table IV: AHSGC has NMI = 0.8587, while EGAE has NMI = 0.8769, so AHSGC is not the best on NMI and the alleged gain is actually a 1.82-percentage-point deficit. The stated 0.73% improvement appears to use NCSC (0.8514) as the second-best method rather than the actual second-best EGAE. The narrative must be corrected to compare against EGAE and to restrict the 'best' claim to the metrics that actually support it.","section":"Section IV-C.1, Table IV"},{"comment":"The claim 'our AHSGC method still performs the best among all investigated clustering methods' on Pavia University is not supported by Table V. AHSGC has NMI = 0.5251, which is lower than both K-means (0.5529) and EGAE (0.5372), and Purity = 0.6980, which is lower than EGAE (0.7126). The conclusion that AHSGC is best across the board is therefore false as written; the empirical summary must identify the specific metrics and datasets for which AHSGC is actually leading.","section":"Section IV-C.2, Table V"},{"comment":"The graph reconstruction loss L_g = ||ZZ^T - A̅||_F^2 trains Z to match A̅, but A̅ is constructed in Eq. (24) from edge sets obtained via Eqs. (20)-(23) using S = ZZ^T and pseudo-labels c_i = argmax_k q_ik computed from the same Z. The reconstruction target is therefore a thresholded function of the model's own output, not an independent supervisory graph. If the high-confidence pseudo-labels are systematically wrong, the update loop can reinforce errors instead of correcting them. This circularity is load-bearing for the homophily-enhanced structure-learning contribution; the authors should either provide an independent reconstruction target or report a controlled experiment (e.g., randomizing or corrupting the pseudo-labels) showing that the loop does not merely chase its own output.","section":"Section III-D, Eqs. (24)-(25)"},{"comment":"The paper states that all methods are run ten times to eliminate bias, but Tables IV-VI report only point estimates with no standard deviations, confidence intervals, or significance tests. Moreover, the hyperparameters (xi, eta, T, L) are selected by grid search on the test set, as described in Section IV-F. Several reported advantages are very small (e.g., Salinas Purity 0.8368 for AHSGC versus 0.8357 for EGAE), so without variance estimates or a validation-based selection protocol these differences are not established. Please report mean and standard deviation over the ten runs and separate validation data from the test set for hyperparameter selection.","section":"Section IV-B.3 and Section IV-F"},{"comment":"The adaptive filter in Eq. (11) depends on a parameter mu that is described as a learnable balance between high-pass and low-pass information, but no update rule, initialization, or constraint for mu is provided. Table III lists seven preset parameters yet omits mu, and the parameter analysis in Section IV-F does not discuss it. Since the 'adaptive filter' is a central contribution, the training or selection of mu must be specified.","section":"Section III-B, Eq. (11)"}],"minor_comments":[{"comment":"The section title 'EXPLEMENTS' should be corrected to 'EXPERIMENTS'.","section":"Section IV heading"},{"comment":"Three distinct figures are all numbered 'Fig. 8': the parameter-sensitivity plots, the t-SNE visualization, and the ablation bar charts. These should be renumbered sequentially.","section":"Figures in Sections IV-F through IV-H"},{"comment":"The notation for Z is inconsistent: Table I defines Z as the pixel-superpixel correlation matrix in R^{hw x N}, while Eq. (13) uses Z for the graph encoder output in R^{N x D}. Please use distinct symbols or clarify the reuse.","section":"Table I and Eq. (13)"},{"comment":"The text refers to eta as the 'inter-cluster edge removal ratio', but Eq. (23) describes it as the percentage of retained edges, and Table I calls it the removal ratio. This ambiguity should be resolved, as the interpretation affects the graph sparsification behavior.","section":"Section III-D, Eq. (23)"},{"comment":"The sentence reporting Salinas results lists four numerical values for five metrics; please report all five values explicitly and consistently with Table IV.","section":"Section IV-C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's own results contradict headline claims in two datasets; this is not a matter of disagreement with external benchmarks but an internal inconsistency in the reported tables and text. The self-training graph-update loop also needs a direct response: without an external or independent target, the reconstruction loss may be a form of self-confirmation rather than a principled objective. I would not recommend rejection solely on these grounds, but the authors need to correct the empirical claims and substantially strengthen the statistical and conceptual support before the paper can be considered acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: AHSGC is a sensible deep graph clustering pipeline for hyperspectral images, and the OA gains over nine baselines on Salinas, PaviaU, and Trento are worth taking seriously. But the paper overstates its own results: Section IV-C.1 claims best clustering performance across five metrics on Salinas, and Table IV shows AHSGC's NMI (0.8587) is below EGAE (0.8769). On PU, AHSGC also trails EGAE on NMI and Purity. The claim 'best clustering performance' is false as written.\n\nWhat's new: the combination of SLIC superpixels, an adaptive low/high-pass graph filter, and homophily-based edge recovery/removal driven by pseudo-labels is not in the prior work I know. The ablation and complexity analyses are useful, and the runtime figures suggest the method is practical. The reported OA (83.60/63.65/86.03) is a real improvement over the compared baselines if it holds.\n\nSoft spots, in order: (1) The reporting inconsistency matters because it erodes confidence in the headline claim. The text and tables need to agree, and the authors should specify which metrics actually improve. (2) No error bars or standard deviations despite ten runs; with random initialization and K-means at the end, variance could be notable. (3) Hyperparameters are selected by grid search on the test sets, so the results are optimistic relative to a fair evaluation. (4) The graph reconstruction loss (Eq. 25) trains Z to match an adjacency built from Z's own pseudo-labels and edge updates; the loop has no external anchor and can entrench mistakes. This is a design concern, not a proven flaw, but it should be discussed and ideally validated with an external graph or a sanity check. (5) Code is not actually released, despite the promise.\n\nFor whom: someone working on deep clustering for hyperspectral images will find the framework and comparisons useful. A general ML reader would not. The paper deserves a serious referee, but it needs a major revision before acceptance: correct the claims, add variance estimates, report a non-test-set hyperparameter selection, and address the self-referential graph update.\n\nRecommendation: send to review, but expect the authors to fix the reporting and provide code.","headline":"Strong OA numbers on three HSI benchmarks, but the paper's own tables contradict its 'best on all metrics' claim, and the graph-update loop is self-referential.","tokens_in":21859,"tokens_out":2024,"would_cite":false,"duration_ms":20121,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adaptive filters plus pseudo-label-driven edge rewiring make the graph itself a trainable object, yielding the best reported clustering accuracy on three hyperspectral benchmarks.","keywords":["hyperspectral image clustering","adaptive graph filter","homophily-enhanced structure learning","self-training clustering","graph edge sparsification","superpixel graph","joint network optimization","unsupervised land-cover mapping"],"falsifier":"Run AHSGC on Salinas with the same number of added and removed edges, but choose the added and removed edge sets at random instead of by confidence-ranked pseudo-labels; if overall accuracy stays near the reported 83.60%, the pseudo-label guidance is not the source of the gain.","tokens_in":20677,"feed_emoji":"🛰️","tokens_out":13962,"duration_ms":124189,"temperature":0.7,"pith_summary":"The paper tries to establish that hyperspectral image clustering with zero labeled pixels improves substantially when the graph is not fixed but is rewritten as the clustering proceeds. The proposed AHSGC method (a homophily structure graph-learning framework with an adaptive filter) segments the image into superpixels, encodes them with an adaptive graph filter that learns to keep both low-frequency similarity signals and high-frequency difference signals, and then sparsifies graph edges from its own confident pseudo-labels: edges are added inside clusters and removed between clusters. It reports the best clustering accuracy among the methods compared on all three benchmarks, with overall accuracies of 83.60% on Salinas, 63.65% on Pavia University, and 86.03% on Trento, while keeping training time lower than several deep clustering baselines. If this holds, land-cover maps for new hyperspectral scenes could be produced without any labeled training samples and with modest compute.","feed_headline":"Self-rewiring graph lifts hyperspectral clustering accuracy to 86%","feed_subtitle":"Updating edges from confident pseudo-labels yields top accuracy on three hyperspectral benchmarks, no labels needed.","key_machinery":"The central object carrying the argument is the homophily-enhanced structure-learning loop coupled with an adaptive graph filter. The adaptive filter, $F_A=\\mu(S_{rw})^k X+(1-\\mu)(I-S_{rw})^k X$, replaces the standard low-pass-only graph convolution with a learnable mixture of random-walk-normalized propagation and its Laplacian complement, so the encoder can preserve both smooth and non-smooth node features. The structure-learning loop uses pseudo-labels and high-confidence node subsets to estimate node similarities in embedding space, then rewrites the graph by intra-cluster edge recovery and inter-cluster edge removal. The rewritten adjacency $\\bar{A}$ enters the reconstruction loss, which is what trains the embeddings toward the task-specific graph; the whole joint objective $\\mathcal{L}_O$ is what makes the graph update and the network training happen in one loop.","core_discovery":"For an unlabeled hyperspectral image, AHSGC constructs a superpixel-level graph and learns node embeddings through a stack of adaptive graph filters, $F_A=\\mu(S_{rw})^k X+(1-\\mu)(I-S_{rw})^k X$, so the encoder can keep both similarity-bearing low-frequency content and difference-bearing high-frequency content. A self-training decoder turns embeddings into soft cluster assignments $q_{ik}$ via a Student's $t$-distribution and sharpens them with KL divergence, giving pseudo-labels $c_i=\\arg\\max_k q_{ik}$. The homophily-enhanced structure-learning module then takes the top-$\\gamma$ most confident nodes per cluster, estimates pairwise similarities $S_{ij}^k=Z_i^k (Z_j^k)^T$ inside those subsets, adds edges among the most similar same-label pairs, removes edges between high-similarity different-label pairs, and updates the adjacency to $\\bar{A}=A-A_{\\mathcal{E}_{rm}}+A_{\\mathcal{E}_{rc}}$. A reconstruction loss $\\mathcal{L}_g=\\|ZZ^T-\\bar{A}\\|_F^2$ trains the embeddings to reproduce this sparsified structure, and the combined objective $\\mathcal{L}_O=\\mathcal{L}_c+\\mathcal{L}_g$ is optimized jointly before K-means on the final embeddings. The central claim is that this feedback loop corrects erroneous graph connections and extracts features that fixed-graph deep clustering methods miss.","pith_inferences":["Because the graph update is driven by the clustering's own pseudo-labels, early confident mistakes are likely to be amplified rather than corrected; injecting controlled label noise in the first iterations and measuring recovery would test this.","The adaptive-filter mixing idea is not tied to hyperspectral images and could be carried to general graph clustering or semi-supervised node classification, where the low- versus high-frequency balance is usually fixed by architecture rather than learned.","The edge-sparsification rules assume homophily, so on heterophilic scenes or graphs they may remove useful cross-class edges; the reported gains may not transfer without retuning the recovery and removal ratios."],"forward_implications":["If the reported numbers hold, new hyperspectral scenes can be clustered into land-cover classes with no labeled pixels, with overall accuracy above 83% on Salinas and 86% on Trento.","The paper's ablations show each module contributes: removing the homogeneous-region generation, the adaptive filter encoder, or the homophily-enhanced structure learning lowers the five metrics on all three datasets.","Because the final step is K-means on embeddings trained to match the sparsified graph, the inference step is cheap; the reported test times are under a tenth of a second on all three datasets.","The adaptive-filter ablation supports the paper's claim that high-frequency boundary and difference information is a useful signal for HSI clustering, not just noise to be filtered out."],"supporting_citations":[{"why":"Provides the superpixel segmentation routine that converts the image into the superpixel-level graph nodes the method starts from.","marker":"[37]"},{"why":"Supplies the homophily-enhanced structure-learning idea that the paper adapts into intra-cluster edge recovery and inter-cluster edge removal.","marker":"[32]"},{"why":"Supplies the adaptive hybrid graph-filter idea that the paper adapts to combine low- and high-frequency node signals.","marker":"[33]"},{"why":"Establishes the low-pass graph-filter HSI clustering approach that this method extends with adaptive high-frequency components.","marker":"[27]"},{"why":"Provides the deep attentional graph autoencoder with self-training that serves as a primary graph-clustering baseline and inspiration.","marker":"[29]"},{"why":"Supports the sharpened auxiliary-distribution design that prevents the self-training clustering from collapsing into one cluster.","marker":"[35]"},{"why":"Provides the superpixel-level contrastive subspace clustering baseline used in the comparison experiments.","marker":"[17]"},{"why":"Supplies the dual graph autoencoder for spectral-spatial HSI clustering that the adaptive-filter contribution is compared against.","marker":"[26]"}],"fun_headline_variants":["Graph rewiring via pseudo-labels sharpens HSI clustering","Adaptive filter graph learns to fix its own edges for HSI","Self-trained graph sparsification lifts HSI cluster accuracy","Unlabeled HSI clusters improved by homophily-guided edge updates","Adaptive homophily graph learning refines HSI clusters"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pseudo-labels produced by the current clustering are accurate enough for the confidence-filtered edge additions and removals to improve the graph, because those same pseudo-labels are the only guide for rewriting the adjacency and there is no external correction in the loop.","fun_headline_variants_meta":{"raw":{"variants":["Graph rewiring via pseudo-labels sharpens HSI clustering","Adaptive filter graph learns to fix its own edges for HSI","Self-trained graph sparsification lifts HSI cluster accuracy","Unlabeled HSI clusters improved by homophily-guided edge updates","Adaptive homophily graph learning refines HSI clusters"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000474,"raw_usage":{"total_tokens":2440,"prompt_tokens":1119,"completion_tokens":1321,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":735,"completion_tokens_details":{"reasoning_tokens":1234}},"tokens_in":735,"tokens_out":1321,"duration_ms":12224,"temperature":1.0,"reasoning_tokens":1234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:25:19.255135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run AHSGC on Salinas with the same number of added and removed edges, but choose the added and removed edge sets at random instead of by confidence-ranked pseudo-labels; if overall accuracy stays near the reported 83.60%, the pseudo-label guidance is not the source of the gain.","supporting_citations":[{"cited_title":"SLIC superpixels compared to state-of-the-art superpixel methods,","cited_arxiv_id":null,"evidence_quote":"Provides the superpixel segmentation routine that converts the image into the superpixel-level graph nodes the method starts from."},{"cited_title":"Homophily -enhanced structure learning for graph clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the homophily-enhanced structure-learning idea that the paper adapts into intra-cluster edge recovery and inter-cluster edge removal."},{"cited_title":"Homophily -Related: Adaptive Hybrid Graph Filter for Multi -View Graph Clustering","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive hybrid graph-filter idea that the paper adapts to combine low- and high-frequency node signals."},{"cited_title":"Self-supervised locality preserving low- pass graph convolutional embedding for large -scale hyperspectral image clustering,","cited_arxiv_id":null,"evidence_quote":"Establishes the low-pass graph-filter HSI clustering approach that this method extends with adaptive high-frequency components."},{"cited_title":"Beyond homophily: Reconstructing structure for graph -agnostic clustering","cited_arxiv_id":null,"evidence_quote":"Supports the sharpened auxiliary-distribution design that prevents the self-training clustering from collapsing into one cluster."},{"cited_title":"Superpixel contracted neighborhood contrastive subspace clustering network for hyperspectral images,","cited_arxiv_id":null,"evidence_quote":"Provides the superpixel-level contrastive subspace clustering baseline used in the comparison experiments."},{"cited_title":"Spectral–spatial feature extraction with dual graph autoencoder for hyperspectral image clustering,","cited_arxiv_id":null,"evidence_quote":"Supplies the dual graph autoencoder for spectral-spatial HSI clustering that the adaptive-filter contribution is compared against."}],"review_version":1}