{"id":"38f1ccc8-e68a-4c7f-8f18-f49f9e6969b5","arxiv_id":"2411.17291","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A self-guided, label-free hyperparameter tuning method for subspace clustering that splits hyperparameter intervals based on agreement between pseudo-labels, reaching within 5% to 7% of oracle-tuned performance.","lead":"This paper proposes a way to tune subspace clustering algorithms without any labeled data: it compares clusterings produced by neighboring hyperparameter values and zooms in on the interval where they agree most. The method typically lands within 5% to 7% of the accuracy of an oracle that uses true labels, and adds a visualization of cluster-specific subspace bases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (14) defines ACC by direct label equality with no permutation alignment; because cluster IDs are arbitrary, the reported pseudo-label ACC is not a valid clustering agreement, undermining ACC-based LFSG results unless the code aligns labels.","rationale":"The reader's weakest assumption emphasized the smoothness/monotonicity of h; I agree that this is unsupported and Figure 2 shows a failure mode. However, the more load-bearing issue is that h itself, as defined in Eq. (14), is not a valid clustering agreement measure unless labels are aligned. Without alignment, ACC between pseudo-labelings is meaningless, so the ACC-based results in the paper cannot be interpreted. This is a concrete, checkable defect rather than a heuristic risk. The empirical pattern (some tables, code availability) suggests the method may still work, especially via NMI, so a conditional verdict remains appropriate: the authors must fix the ACC definition or confirm that the code aligns labels, and re-run affected experiments. I would not reject outright because the NMI-based results and code could support the central claim after correction.","tokens_in":30995,"tokens_out":4886,"duration_ms":49760,"concrete_test":"Inspect the public MATLAB implementation (https://github.com/ikopriva/LFSGSC) for any label-alignment step before Eq. (14). Then re-run the LSR SC experiment on MNIST and USPS (Table 2) computing pseudo-label ACC with Hungarian alignment and with the paper's direct equality; compare the selected lambda and the final clustering ACC. If the selected lambda or the reported LFSG-ACC values change materially, Eq. (14) is mis-specified and ACC-based claims need revision. Also rerun Figure 1(b) with aligned ACC to see whether the pseudo-label ACC curve still peaks in the same interval.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (14) in Section 3.1 defines ACC between pseudo-labelings y_i and y_{i+1} as the fraction of samples with identical label values, with no Hungarian or other alignment step. Spectral clustering followed by k-means returns cluster IDs only up to an arbitrary permutation, so this quantity is not the standard clustering accuracy and can be near zero even when the two partitions are identical up to relabeling. All LFSG-ACC results in Tables 2-6, 8, and 9 are therefore computed from a metric that the paper never defines correctly; the selected hyperparameter intervals may reflect label-order artifacts rather than true agreement. The paper's central claim (typically 5-7% lower than oracle) rests on these ACC selections as much as on NMI. Figure 2 also documents a separate failure of the smoothness/monotonicity assumption, but the label-alignment defect is more fundamental because it affects the validity of the metric itself. If the public code applies an implicit alignment (e.g., same k-means initialization or label matching), that needs to be stated and the equation corrected; otherwise the ACC-based experiments are not reproducible as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a label-free self-guided hyperparameter optimization (LFSG-HPO) method for subspace clustering. For a chosen SC algorithm and an initial hyperparameter grid, the method computes ACC or NMI between pseudo-labelings produced by the algorithm at neighboring hyperparameter values, assumes these agreement metrics to be smooth and locally monotone in the hyperparameter, selects the interval where the agreement is maximal, and iteratively refines that interval by splitting it into halves or thirds until a relative-error criterion is met. The same idea is extended to two-hyperparameter algorithms by fixing one parameter while optimizing the other. The paper also contributes an out-of-sample extension for kernel LSR SC and a visualization method for interpreting clusters via estimated subspace bases. Experiments compare oracle-tuned and LFSG-tuned versions of LSR, kernel LSR, graph-filtering LSR, SSC, S0L0 LRSSC, LMVSC, and MLME on six single-view and three multi-view datasets, claiming a typical performance loss of 5% to 7% relative to the oracle.","tokens_in":31286,"tokens_out":5786,"duration_ms":53352,"significance":"If the central assumptions hold, this is a practically useful contribution: it offers a label-free, general-purpose tuning strategy for existing subspace clustering algorithms, with released code and a relatively simple algorithmic core. The paper contains a broad experimental study with statistical testing across multiple algorithms and datasets, and the interpretability component via subspace-base visualization is a nice addition. However, the contribution rests on an unproven smoothness/monotonicity premise that is shown to fail in Figure 2, and the ACC metric in Eq. (14) is not valid as written because it does not account for the arbitrary permutation of cluster IDs. These issues affect the validity of a large part of the reported ACC-based experiments, so the paper is not acceptable in its current form.","major_comments":[{"comment":"The accuracy between two pseudo-labelings y_i and y_{i+1} is defined as the fraction of samples for which the labels are exactly equal, with no label-alignment step. Since the SC pipeline (spectral embedding followed by k-means) returns cluster IDs only up to an arbitrary permutation, this quantity is not a valid clustering accuracy: two identical partitions can have near-zero ACC under Eq. (14), while two very different partitions can agree by chance. All ACC-based LFSG selections in Tables 2-6, 8, and 9 therefore rest on an undefined metric unless the implementation silently aligns cluster IDs (e.g., via Hungarian matching or by reusing the same k-means initialization). The authors must either correct Eq. (14) to include a proper alignment step, or explicitly document and justify the alignment used in the provided code.","section":"Section 3.1, Eq. (14)"},{"comment":"The method's load-bearing premise is that h(y_i,y_{i+1}), i.e., ACC or NMI between pseudo-labels at neighboring hyperparameters, is a smooth and locally monotone function whose maximizing subinterval contains the hyperparameter that maximizes true clustering performance. No proof or theoretical characterization is given, and Figure 2 documents a concrete failure when the initial grid is too dense: the maximal pseudo-label agreement occurs on an interval far from the true optimal lambda. Section 4.3 admits that the method 'critically depends' on the smoothness assumption and on the quality of the initial hyperparameter search space. The paper needs either a formal condition under which the selection rule is correct, or an empirical demonstration that the Figure 2 failure mode is rare and detectable without oracle labels. As written, the claimed 'typically 5% to 7% lower than oracle' is not guaranteed by the experiments.","section":"Section 3.1, Eqs. (16)-(18), Figure 2, Section 4.3"},{"comment":"The abstract's claim that the proposed method 'typically achieves clustering performance that is 5% to 7% lower than that of the oracle versions' is not supported by all of the reported results. For example, in Table 6 on ORL, the LFSG-ACC in-sample accuracy is 57.76% versus the oracle-ACC value of 71.03%, a gap of about 13 percentage points, and the F1-score gap is about 19 percentage points. Smaller but still larger-than-claimed gaps appear elsewhere, e.g., Table 5 EYaleB F1-score. If 'typically' means 'in most but not all cases' or refers to a median, that should be stated explicitly with a summary of the full distribution of performance gaps; otherwise the abstract overstates the method's consistency.","section":"Abstract, Section 4.1.5, Table 6"},{"comment":"Table 10, which should report the clustering performance of the parameter-free FPMVS-CAG algorithm, is missing from the manuscript: only its caption appears before Section 4.3. The text in Section 4.2.1 asserts that LFSG LMVSC outperforms FPMVS-CAG on two of three datasets, but without the actual numbers in Table 10 this comparison is unverifiable. The authors should complete Table 10 with the same metrics and experimental conditions used for Tables 8 and 9, or remove the comparative claim.","section":"Section 4.2.3, Table 10"}],"minor_comments":[{"comment":"Step 7 refers to 'the relative error criterion (13)', but the stopping criterion is defined in Eq. (21), not Eq. (13). Please correct the cross-reference.","section":"Algorithm 2, step 7"},{"comment":"There is an inconsistency between the text and the Figure 2 caption: the text says the grid values lambda_1=10^-7 and lambda_2=10^-6 are 'set too close', while the caption states that the maximal pseudo-label agreement occurs between lambda_2=10^-6 and lambda_3=10^-5 and that the new iteration starts at those borders. Please clarify which neighboring pair is responsible for the failure.","section":"Figure 2 caption and Section 3.1 text"},{"comment":"For the ORL dataset, the LFSG-NMI row reports exactly the same values as the ORACLE-NMI row, including identical hyperparameters and identical performance metrics. This looks like a copy-paste error and should be corrected, since it contradicts the surrounding text reporting a large performance drop on ORL.","section":"Section 4.1.5, Table 6, ORL row"},{"comment":"Equation (33) is used twice: once in Section 3.5 for the SVD of each cluster partition, and again in Section 4.1.4 for the robust SSC formulation. Please renumber one of these equations.","section":"Section 4.1.4, Eq. (33)"},{"comment":"The manuscript contains several typographical and wording issues, including 'obatined' for 'obtained', 'prehistoric evaluations' for what should presumably be 'prior historical evaluations', and inconsistent spacing around mathematical expressions. A careful proofreading pass is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The label-alignment issue in Eq. (14) is the most consequential concern: if the released code already aligns cluster labels (for example by using a fixed initialization or Hungarian assignment), the authors can address the issue by documenting that step and correcting the equation. The missing Table 10 and the failure mode in Figure 2 should also be resolved in revision. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the LFSG subspace clustering HPO paper. The core idea is simple and useful: on a coarse hyperparameter grid, run the SC algorithm, compute NMI (or ACC) between pseudo-labelings at neighboring hyperparameter values, pick the interval with the highest agreement, split it into thirds, and repeat until the interval is narrow. No labeled data, no internal clustering validity index, no pretext task. That combination for SC hyperparameter selection appears to be new, and the paper shows it consistently lands within a few percent of oracle-tuned performance across six datasets and five SC algorithms, with public MATLAB code and reproducible data. The graph-filtering LSR formulation and out-of-sample kernel extension are side contributions that are fine.\n\nThe main contribution is the empirical evidence, and it is honest: the paper reports cases where LFSG underperforms oracle by more than 10% (e.g., S0L0 on ORL, kernel LSR on EYaleB), and the Wilcoxon tests show many differences are insignificant. The 5-7% headline claim is supported by most tables. The paper also explicitly acknowledges in Section 4.3 that the smoothness assumption is a limitation and that the initial search space matters. Good.\n\nNow the soft spots, in proportion.\n\nThe biggest technical problem is Eq. (14). ACC is defined as the fraction of samples with identical label values between two pseudo-labelings, with no permutation alignment. Cluster IDs from spectral clustering plus k-means are arbitrary up to relabeling. As written, this ACC is not a valid clustering agreement and can be near zero for identical partitions. NMI is permutation-invariant and fine, but many of the ACC-based experiments rest on an undefined metric. Either the code implicitly aligns labels (then state it in the equation) or the ACC selection results need to be redone. This is a load-bearing blemish, not just cosmetic.\n\nThe smoothness/monotonicity assumption is unproven, and Figure 2 shows a concrete failure when the grid is too dense. The paper admits this. It means the method is a heuristic with a known failure mode, not a principled optimizer. For many real grids it works, but there is no characterization of when it fails beyond the example.\n\nThe FPMVS-CAG comparison in Table 8/10 is a bit awkwardly placed, but the claimed pattern (better on two datasets, worse on one) matches the table, so it is not misreported.\n\nWho is this for? People who need to tune subspace clustering on unlabeled data and are willing to accept a heuristic. It deserves a serious referee, but the ACC equation needs correcting, and the authors should either prove a limited version of the smoothness condition or state more precisely the empirical conditions under which the interval-splitting rule is sound. I would send it out for review.","headline":"A practical pseudo-label smoothness heuristic for tuning subspace clustering without labels; the ACC definition in Eq. (14) is broken without label alignment, but the NMI-based results likely survive.","tokens_in":114,"tokens_out":2220,"would_cite":true,"duration_ms":53633,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","68T05","68T10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Subspace clustering can be tuned with no labeled data, the paper claims, by comparing pseudo-labels across neighboring hyperparameter values.","keywords":["subspace clustering","hyperparameter optimization","label-free learning","pseudo-labels","self-guided tuning","normalized mutual information","graph filtering","interpretability"],"falsifier":"Run the interval-refinement on COIL20 with the dense grid from Figure 2 (spacing from $10^{-7}$ to $10^{-6}$): pseudo-label ACC maximizes on [$10^{-6}$, $10^{-5}$], while oracle accuracy peaks at $\\lambda$ = 0.1, so if the method returns a subinterval with near-chance true accuracy, the smoothness-to-optimality premise is refuted; a coarser grid that lands on the true optimum would only confirm the premise conditionally.","tokens_in":30777,"feed_emoji":"🎛️","tokens_out":5824,"duration_ms":51074,"temperature":0.7,"pith_summary":"Hyperparameter tuning of subspace clustering normally requires labeled validation data. This paper claims that tuning can be done label-free: run the clustering algorithm over a grid of hyperparameter values, and measure how similar the resulting pseudo-labelings are at neighboring values, using accuracy (ACC) or normalized mutual information (NMI). If those agreement curves are smooth, the interval where the agreement is highest should contain the hyperparameter that would maximize true clustering accuracy. The paper tests this on six datasets with linear, kernel, graph-filtered, and multi-view subspace clustering algorithms, and reports performance typically 5% to 7% below the oracle (label-tuned) versions. The same surrounding machinery also gives an out-of-sample extension and a way to visualize learned subspace bases for interpretability.","feed_headline":"Tune subspace clustering with zero labeled data","feed_subtitle":"Pseudo-label agreement across neighboring settings finds near-oracle hyperparameters, 5-7 percent below oracle accuracy.","key_machinery":"The engine is the neighbor-pair agreement function h(y_i, y_{i+1}), e.g. ACC (Eq. 14) or NMI (Eq. 15), computed between pseudo-labelings generated by the same subspace clustering algorithm at consecutive grid values lambda_i and lambda_{i+1}. Eq. (18) locates the subinterval where this agreement is maximal; Eqs. (16)-(17) assert that it is locally monotone; Eq. (20) refines the interval by thirds; Eq. (21) stops when the relative error falls below epsilon. For two hyperparameters the search alternates: fix tau, find lambda*, then fix lambda*, find tau*.","core_discovery":"The paper's central claim is that external labels are unnecessary for hyperparameter selection in subspace clustering: the ACC or NMI between clusterings produced at adjacent hyperparameter values acts as a proxy for true clustering quality. Under a smoothness and monotonicity assumption on this proxy, the globally best hyperparameter lies inside the subinterval where neighbor-agreement is maximal, and repeated splitting of that interval into thirds converges to it. With relative-error stopping at epsilon = 0.001, the method selects a hyperparameter whose resulting clustering is typically 5% to 7% worse (in ACC, NMI, or F1) than the oracle's, sometimes statistically indistinguishable, and on out-of-sample data often within 1%. The paper further extends the same self-guided selection to two-parameter algorithms by alternating one-parameter searches, and to out-of-sample points by fitting subspace bases to the in-sample partitions and assigning test points to the nearest subspace.","pith_inferences":["Because the pseudo-label agreement only needs the clustering algorithm's outputs, the same interval-refinement machinery could be dropped into any clustering method with a continuous hyperparameter, not just subspace clustering.","A testable extension is to align cluster indices (for example, by Hungarian matching) before computing ACC between neighboring pseudo-labelings; spectral clustering's label-permutation invariance would otherwise make the equality in Eq. (14) sensitive to the k-means initialization.","The dense-grid failure in Figure 2 suggests a diagnostic rule: run the method on a coarse grid first and check that the chosen interval's pseudo-label agreement is not near 100%, which would signal that adjacent labelings are trivially identical; the paper's own visualization step could be used for that check.","An even simpler variant that compares pseudo-labels at only three points per iteration instead of four is noted in the paper and would reduce runtime, with the paper reporting that it affects clustering performance minimally."],"forward_implications":["Any existing subspace clustering algorithm with tunable hyperparameters can be reused in label-poor domains like medicine without a pretext task or an internal cluster-quality index.","The performance gap to oracle tuning is typically 5% to 7%, and often statistically insignificant on out-of-sample data, so the practical cost of removing labels is bounded.","The graph-filtering and kernel out-of-sample formulations (Algorithm 1 and Section 3.4) widen the set of subspace clustering algorithms that can benefit, including nonlinear data.","The subspace-basis visualization gives domain experts a way to judge clustering quality and refine the initial search space before re-running the optimizer."],"supporting_citations":[{"why":"Supplies the LSR SC algorithm used as the running example and the analytic lambda-dependent solution (Eq. 6) that the pseudo-label grid is built on.","marker":"[17]"},{"why":"Introduces the smoothness assumption on performance metrics over hyperparameters that the proposed method adopts and operationalizes without historical evaluations.","marker":"[52]"},{"why":"Documents why standard internal clustering quality metrics fail for union-of-linear-subspaces models, motivating pseudo-label ACC/NMI as the selection signal.","marker":"[43]"},{"why":"Defines the SSC algorithm, one of the five single-view algorithms evaluated with the proposed label-free HPO.","marker":"[15]"},{"why":"Defines the S0L0 LRSSC algorithm, another evaluation subject, including the ADMM-based solver whose penalty constant becomes a second hyperparameter.","marker":"[16]"},{"why":"Defines the LMVSC multi-view algorithm with anchors M and lambda, used in the multi-view experiments.","marker":"[24]"},{"why":"Defines the MCMLE/MLME multi-view algorithm with hyperparameters alpha and beta, used in the multi-view experiments.","marker":"[72]"},{"why":"Provides the graph-filtering alternative to kernel methods that the paper turns into Algorithm 1's graph-filtered LSR SC.","marker":"[68]"},{"why":"Supplies the subspace-basis estimation and point-to-subspace distance equations (Eqs. 23-25) used for out-of-sample clustering.","marker":"[37]"},{"why":"Provides the hyperparameter-free FPMVS-CAG baseline that the LFSG versions are compared against on multi-view data.","marker":"[20]"}],"fun_headline_variants":["Self-guided subspace clustering: no labels, near-oracle tuning","Pseudo-labels self-tune subspace clustering to near-oracle","Label-free tuning for subspace clustering, oracle-close","Zero-label tuning for subspace clustering: oracle-close","No labels? Self-tune subspace clustering via pseudo-agreement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ACC or NMI between pseudo-labelings at neighboring hyperparameters varies smoothly and monotonically, so the subinterval with the highest neighbor agreement always contains the hyperparameter that would maximize true clustering accuracy; the paper does not prove this, and its own Figure 2 shows a case where a too-dense grid violates it.","fun_headline_variants_meta":{"raw":{"variants":["Self-guided subspace clustering: no labels, near-oracle tuning","Pseudo-labels self-tune subspace clustering to near-oracle","Label-free tuning for subspace clustering, oracle-close","Zero-label tuning for subspace clustering: oracle-close","No labels? Self-tune subspace clustering via pseudo-agreement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001151,"raw_usage":{"total_tokens":4818,"prompt_tokens":1038,"completion_tokens":3780,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":3697}},"tokens_in":654,"tokens_out":3780,"duration_ms":27178,"temperature":1.0,"reasoning_tokens":3697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:16:25.122627+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the interval-refinement on COIL20 with the dense grid from Figure 2 (spacing from $10^{-7}$ to $10^{-6}$): pseudo-label ACC maximizes on [$10^{-6}$, $10^{-5}$], while oracle accuracy peaks at $\\lambda$ = 0.1, so if the method returns a subinterval with near-chance true accuracy, the smoothness-to-optimality premise is refuted; a coarser grid that lands on the true optimum would only confirm the premise conditionally.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LSR SC algorithm used as the running example and the analytic lambda-dependent solution (Eq. 6) that the pseudo-label grid is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the smoothness assumption on performance metrics over hyperparameters that the proposed method adopts and operationalizes without historical evaluations."},{"cited_title":"Lipor, L","cited_arxiv_id":null,"evidence_quote":"Documents why standard internal clustering quality metrics fail for union-of-linear-subspaces models, motivating pseudo-label ACC/NMI as the selection signal."},{"cited_title":"Elhamifar, R","cited_arxiv_id":null,"evidence_quote":"Defines the SSC algorithm, one of the five single-view algorithms evaluated with the proposed label-free HPO."},{"cited_title":"Brbić, I","cited_arxiv_id":null,"evidence_quote":"Defines the S0L0 LRSSC algorithm, another evaluation subject, including the ADMM-based solver whose penalty constant becomes a second hyperparameter."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the LMVSC multi-view algorithm with anchors M and lambda, used in the multi-view experiments."},{"cited_title":"Zhong, C","cited_arxiv_id":null,"evidence_quote":"Defines the MCMLE/MLME multi-view algorithm with hyperparameters alpha and beta, used in the multi-view experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the graph-filtering alternative to kernel methods that the paper turns into Algorithm 1's graph-filtered LSR SC."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the subspace-basis estimation and point-to-subspace distance equations (Eqs. 23-25) used for out-of-sample clustering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the hyperparameter-free FPMVS-CAG baseline that the LFSG versions are compared against on multi-view data."}],"review_version":1}