{"id":"38439448-631e-4411-9f26-b80ffde34ae4","arxiv_id":"2501.04562","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Spherical Double K-Means simultaneously clusters terms and documents using a cosine-style objective, applied here to US inaugural addresses.","lead":"This paper introduces Spherical Double K-Means, an algorithm that clusters the words and the documents of a text corpus at the same time. It applies the method to US presidential inaugural addresses and reports that the resulting clusters follow major historical periods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SDKM's Algorithm 1 does not optimize its stated cosine objective: assignments use unnormalized inner products and the centroid update is the DKM least-squares formula, so the central 'spherical' claim and monotonicity guarantee are unsupported.","rationale":"The reader's weakest assumption identifies exactly the algorithm-objective mismatch: assignment uses unnormalized inner products and the centroid update is asserted to maximize a normalized objective without proof. My reading confirms this is the most load-bearing concern. If the algorithm does not optimize the cosine objective, then the method is not 'spherical' in any operational sense, the monotonicity guarantee in Appendix A is irrelevant to the stated objective, and the application results are produced by an unnormalized inner-product variant of DKM rather than by cosine-based SDKM. The paper's own Section 7 limitation statement about normalization reinforces the concern. The cluster-count selection and computational-efficiency claims are additional weaknesses, but the identity of the algorithm is the foundational issue. Therefore the reader's REJECT verdict remains appropriate, and no adjustment is needed.","tokens_in":18189,"tokens_out":4822,"duration_ms":49177,"concrete_test":"Re-run Algorithm 1 on the inaugural TF-IDF matrix and compare it with a cosine-faithful variant: assign rows by maximizing (x_i/||x_i||) dot (row_k(YV')/||row_k(YV')||), columns by maximizing (x_j/||x_j||) dot (col_q(UY)/||col_q(UY)||), and update Y by row-normalized (V'X'U) rows (without (U'U)^{-1} or (V'V)^{-1} scalings). For 100 random starts, record the normalized objective f = tr(X'UYV')/sqrt(tr(X'X)tr(UYV'VY'U')) at each iteration. If the paper's algorithm fails to show monotone increase of f or converges to a lower f than the cosine variant, the central claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SDKM extends SKM by using cosine similarity in co-clustering, but the algorithm does not implement cosine geometry. In Algorithm 1, Step 1 assigns row i to cluster k by maximizing x_i(YV')I_k, an unnormalized dot product, and Step 3 assigns column j similarly via x'_j(UY)W_q. No row normalization of X appears, and the application uses TF-IDF weights without row normalization; Section 3.1 only normalizes the centroid matrix Y by rows, and Section 7 itself concedes the normalization issue is open. The centroid updates in Steps 2 and 4 are the normalized DKM least-squares updates Y = (U'U)^{-1}U'XV(V'V)^{-1}/||...||, which solve min ||X - UYV'||^2 in the unconstrained case, not the maximization of the stated cosine objective. For fixed U,V, maximizing tr(X'UYV') subject to row-unit-norm Y gives Y_{k*} proportional to the k-th column of V'X'U, without the cluster-size scalings (U'U)^{-1} and (V'V)^{-1}. Moreover, Section 3 identifies the matrix cosine similarity with tr(X'Xt), omitting the denominator; the normalized objective is introduced later but never used in the algorithm or in Appendix A. Appendix A proves monotonicity only for the unnormalized trace, and it relies on the false premise that each substep maximizes that trace. Thus the method is unnormalized inner-product co-clustering, the convergence guarantee does not apply to the stated objective, and the claimed advantages over DKM are not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Spherical Double K-Means (SDKM), a co-clustering method intended to combine Double K-Means (DKM) with Spherical K-Means (SKM) in order to simultaneously partition terms and documents using cosine similarity. The method is formalized as a low-rank approximation X ≈ U Y V' with a cosine-based objective, and Algorithm 1 alternates between updating row memberships U, the centroid matrix Y, column memberships V, and Y again. The authors report simulation studies of cluster recovery and of the pseudo-F index for selecting K and Q, then apply SDKM to US presidential inaugural addresses. The application yields two document clusters (roughly pre- and post-Civil War) and three word clusters, which the authors interpret historically. A comparison with DKM shows a largely similar document partition except for Lincoln's 1865 address. The paper claims that SDKM inherits SKM's benefits---robustness to high dimensionality, sparsity, and noise---and provides a monotonicity proof in Appendix A.","tokens_in":18547,"tokens_out":3940,"duration_ms":38979,"significance":"If the paper's central claim were correct, SDKM would be a useful extension of DKM to cosine geometry for text co-clustering, and the application would provide a suggestive demonstration. The paper also ships simulation code (MATLAB, per Section 4.1) and reports detailed simulation results, which is a strength. However, the central technical claim is not sound: Algorithm 1 does not actually optimize the stated cosine objective, and the monotonicity proof in Appendix A does not apply to the normalized objective introduced in Section 3. The methodological contribution is therefore not established, and the application, while historically interesting, cannot compensate for the mismatch between the proposed method and its implementation.","major_comments":[{"comment":"The algorithm's assignment steps are not cosine assignments. Step 1 assigns row i to cluster k by maximizing x_i (Y V') I_k, and Step 3 assigns column j by maximizing x'_j (U Y) W_q; these are unnormalized inner products. The data matrix in the application is the TF-IDF matrix described in Section 5, and no row normalization of X is performed. Section 3.1 explicitly states that only the centroid matrix Y is normalized by rows, and Section 7 lists simultaneous normalization of rows and columns as an open issue. Therefore Algorithm 1 does not implement cosine similarity despite the paper's repeated claims that SDKM is based on cosine geometry, and the claimed advantages over DKM are unsupported.","section":"Section 3, Algorithm 1"},{"comment":"The centroid update Y = (U'U)^-1 U' X V (V'V)^-1 / ||(U'U)^-1 U' X V (V'V)^-1|| is the DKM least-squares formula (cf. Equation (2) in Section 2.2.1), not the maximizer of the stated cosine objective. For fixed U and V, maximizing tr(X' U Y V') over row-unit-norm Y gives, for each row k, Y_k proportional to the k-th column of V' X' U, without the cluster-size scaling matrices (U'U)^-1 and (V'V)^-1. Thus the claim in Section 3 that this Y maximizes the cosine of the angle between rows of X and X_t is not supported by the derivation, which does not account for the normalization constraint correctly.","section":"Section 3, centroid update"},{"comment":"The monotonicity proof in Appendix A only treats the unnormalized trace f(U,Y,V) = tr(X' U Y V'). The normalized objective introduced in Section 3, f(U,Y,V) = tr(X' U Y V') / sqrt(tr(X'X) tr(U Y V' V Y' U')), is never used in the proof or in the algorithm. Moreover, the proof's inequalities (A.1)–(A.4) assume that each substep maximizes the trace objective; as noted in the two previous comments, Steps 1 and 3 maximize inner products rather than cosine similarities, and Steps 2 and 4 use a least-squares centroid update that does not maximize the trace under the stated normalization. Consequently, the claimed monotonic convergence guarantee does not hold for the stated objective, and the statement in Section 3 that 'the algorithm guarantees that the objective function's value increases monotonically' is unsupported.","section":"Appendix A and Section 3 (normalized objective)"},{"comment":"The selection of K=3 and Q=2 is circular as presented. Table 4 shows the highest pseudo-F value at K=2, Q=2; the authors reject this in favor of K=3, Q=2 because the latter is more interpretable, and then use the interpretability of the resulting clusters (Sections 5.1.1 and 5.1.2, Figures 4–7) as evidence of the method's success. Additionally, the clustergram of a K=10, Q=10 solution (Figures 2 and 3) is used to conclude that 'K=3 distinct clusters of words and Q=2 distinct clusters of documents' emerge, but no explanation is given for why a hierarchical clustering of the 10x10 centroid matrix should determine the true number of clusters in the original data. These issues weaken the empirical evidence for the method's practical value.","section":"Section 5.1, cluster selection"}],"minor_comments":[{"comment":"The denominator of the cosine dissimilarity is written '|b∥' in several places; it should be '∥a∥∥b∥'.","section":"Section 2.2.2, Equation (3)"},{"comment":"The matrix X is used for both the data and the centroid matrix in the derivation of SKM (e.g., 'the centroids matrix X'), which makes the row-normalization arguments hard to follow.","section":"Section 2.2.2, notation"},{"comment":"The sentence 'Here, I and W denote the identity matrices of dimensions K and Q' is confusing because Step 1 uses I_k and Step 3 uses W_q; W is also used as a diagonal norm matrix in Section 2.2.2, so the notation should be clarified.","section":"Section 3, Algorithm 1"},{"comment":"The TF-IDF formula x_ij = n_ij / n_j log10(M/m) does not define n_j; presumably it is the total token count in document j, but this should be stated explicitly.","section":"Section 5, TF-IDF formula"},{"comment":"The text states that RndStarts takes values in [1, 5, 20, 30, 40, 50, 70, 100], but Table 2 also includes 10; the list in the text should match the table.","section":"Section 4.4, Table 2"},{"comment":"The method is called 'Spherical Double K-Means' in the title and abstract but 'Double Spherical K-means' in Algorithm 1; the terminology should be made consistent.","section":"Title and Abstract"}],"recommendation":"reject","confidential_remarks":"The manuscript's central methodological claim---that SDKM optimizes a cosine similarity objective---is contradicted by the algorithm it presents. This is not a local presentation issue: the assignments, the centroid update, and the convergence proof all target a different (least-squares/trace) objective. Even a major revision would require redefining the algorithm to actually implement cosine geometry, or substantially revising the claims and proof, which is beyond a routine revision. The application section is interesting but does not rescue the methodological contribution. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the reader's report is right, and the stress-test note holds up when you check the algebra. The paper is the first to combine Dhillon-Modha spherical k-means with Vichi's double k-means, and the application to presidential inaugural addresses is clearly written, with public data, simulations, and a direct DKM/SDKM comparison. That is real work and worth acknowledging.\n\nThe soft spots are load-bearing. The stated objective is normalized cosine similarity, but the algorithm maximizes an unnormalized trace. Step 1 assigns rows by x_i(YV')I_k and Step 3 assigns columns by x'_j(UY)W_q, with no row normalization of X anywhere. The centroid update is the least-squares DKM formula (U'U)^-1 U'XV (V'V)^-1, renormalized. For fixed U and V, the maximizer of tr(X'UYV') under row-unit-norm Y is proportional to U'XV, without the cluster-size scalings. Those scalings come from minimizing ||X - UYV'||^2, which is a different objective. So the monotonicity proof in Appendix A fails at Step 2, and none of the steps is guaranteed to increase the normalized cosine objective, whose denominator depends on U and Y. The paper's own Section 3.1 concedes that simultaneously normalizing both dimensions is an open issue.\n\nThe cluster-count selection is also circular: pseudo-F points to K=2, Q=2, the authors reject that for interpretability, then use interpretability as evidence of success. And the claimed computational advantage over SKM is never benchmarked; Section 6 even says timings are negligible for both methods.\n\nThis is not a hopeless paper. Recast as inner-product double k-means with unit-norm centroids, or fix the updates to actually maximize the cosine objective, and the empirical parts could stand. As written, the central claim about spherical or cosine co-clustering is unsupported. I would send it to a competent referee rather than desk-reject, because the combination is real and the application is reproducible, but I would not cite it in its current form.","headline":"The reader's report is right: SDKM is a genuinely new combination and the case study is clearly presented, but Algorithm 1 does not maximize the cosine objective the paper claims, so the central theoretical and convergence arguments do not hold as written.","tokens_in":19048,"tokens_out":4297,"would_cite":false,"duration_ms":47125,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","62-08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces Spherical Double K-Means (SDKM), a co-clustering method that partitions documents and terms simultaneously by maximizing the cosine similarity between a term-document matrix and its block-centroid approximation.","keywords":["Spherical double k-means","Co-clustering","Cosine similarity","Term-document matrix","Topic modeling","Text clustering","TF-IDF","Presidential inaugural addresses"],"falsifier":"Run SDKM twice on the same TF-IDF matrix, once exactly as Algorithm 1 prints it and once with every row and every column normalized to unit length before the iterations. If the two runs produce different cluster assignments or objective trajectories, the inner-product steps in Algorithm 1 are not equivalent to cosine similarity, and the paper's central premise fails.","tokens_in":41,"feed_emoji":"📊","tokens_out":9009,"duration_ms":147048,"temperature":0.7,"pith_summary":"The paper proposes Spherical Double K-Means (SDKM), an algorithm that clusters the rows (terms) and columns (documents) of a term-document matrix at the same time, using cosine similarity instead of Euclidean distance. It combines double k-means, which approximates a data matrix by block centroids, with spherical k-means, whose centroids are normalized so that similarity is measured by angle. The authors claim this handles high-dimensional, sparse, noisy text better than standard k-means and gives interpretable word and document clusters. On US presidential inaugural addresses, SDKM splits speeches into an early and a later group and separates vocabulary into three themes. A comparison with double k-means shows nearly the same document partition, except Lincoln's 1865 address, which cosine geometry places with modern speeches.","feed_headline":"Cosine co-clustering sorts inaugural speeches into two eras","feed_subtitle":"SDKM clusters speeches and their words together; Lincoln's 1865 address joins the modern group.","key_machinery":"The load-bearing object is the row-normalized block-centroid matrix $$Y=\\frac{(U'U)^{-1}U'XV(V'V)^{-1}}{\\|(U'U)^{-1}U'XV(V'V)^{-1}\\|},$$ updated inside an alternating scheme that also reassigns term memberships $U$ and document memberships $V$. The objective being maximized is the normalized trace of $X' U Y V'$, a cosine-like similarity between the observed term-document matrix and its block-centroid reconstruction $U Y V'$. This centroid matrix, normalized by rows but not by columns, is what carries the spherical part of SDKM, and the paper's monotonicity proof rests on each of the four update steps increasing or preserving the same trace objective.","core_discovery":"SDKM is a co-clustering method built on the double k-means model $X = U Y V' + E$, where $U$ and $V$ are binary membership matrices for terms and documents and $Y$ is a $K \\times Q$ matrix of block centroids. The paper maximizes the normalized trace objective $f(U,V,Y)=\\frac{\\mathrm{tr}(X' U Y V')}{\\sqrt{\\mathrm{tr}(X'X)\\,\\mathrm{tr}(U Y V' V Y' U')}}$ and claims that this quantity is at most 1, reaching 1 only when the reconstructed matrix $U Y V'$ equals $X$. The algorithm cycles through four updates: assign each term to the cluster whose combined profile has the largest inner product, recompute $Y$ as a row-normalized block average, assign each document similarly, and recompute $Y$ again. The authors argue that these inner-product assignments are cosine comparisons after normalization, and Appendix A proves that each sub-step cannot decrease the objective. On the inaugural-address corpus the method yields $K=3$ term clusters and $Q=2$ document clusters, with the document split falling between 1861 and 1865 and Lincoln's 1865 address placed with the modern group.","pith_inferences":["If the normalization gap is closed by explicitly normalizing document vectors before the inner-product assignments, SDKM would become a literal spherical co-clustering method; a symmetric column-normalized variant could be tested against the current row-normalized version on the inaugural corpus.","Because SDKM's objective is a normalized trace, it is close in spirit to low-rank approximation with angle-based fit; comparing it against a topic model on the same data using held-out coherence would show whether the two-way partition adds information beyond LDA.","The DKM-versus-SDKM discrepancy on Lincoln's 1865 speech is a natural probe: synthetic corpora with intermediate vocabulary proportions should switch clusters under SDKM but not under DKM, offering a controlled test of the cosine mechanism.","A fuzzy or soft-membership version, which the authors name as future work, would turn the hard binary assignments into degrees of membership and could be obtained by replacing the arg-max updates with normalized similarities."],"forward_implications":["Users get a two-way partition: each document cluster is directly characterized by the term cluster(s) it draws on, so topic and keyword extraction happen in the same run.","Cosine geometry has a visible effect on borderline documents: Lincoln's 1865 inaugural moves to the modern cluster under SDKM, illustrating that relative word proportions, not raw frequency, drive assignment.","The pseudo-F index can select K and Q automatically, but it undercounts clusters as noise grows; in the application the authors chose K=3, Q=2 for interpretability even though the index's top value was K=2, Q=2.","Twenty random starts were enough in the authors' simulations to avoid local maxima, so SDKM can be used with that restart policy in practice.","The objective is monotonically non-decreasing across iterations, giving a well-defined stopping rule and a convergence guarantee to a local maximum."],"supporting_citations":[{"why":"Introduces double k-means, the $X = U Y V' + E$ block model and alternating update scheme that SDKM inherits.","marker":"(Vichi, 2001)"},{"why":"Defines spherical k-means for sparse text with cosine similarity and normalized centroids, the row-side ingredient of SDKM.","marker":"(Dhillon & Modha, 2001)"},{"why":"Gives the cosine-dissimilarity formulation and the normalized objective that SDKM adapts to the co-clustering setting.","marker":"(Hornik et al., 2012)"},{"why":"Supplies the pseudo-F index used to choose the numbers of row and column clusters.","marker":"(Rocci & Vichi, 2008)"},{"why":"Provides the between/within-cluster variance ratio on which the pseudo-F index is built.","marker":"(Caliński & Harabasz, 1974)"},{"why":"Provides the quanteda package and the US presidential inaugural corpus used in the application.","marker":"(Benoit et al., 2018)"},{"why":"Provides the Adjusted Rand Index used to evaluate partition recovery in the simulations and the DKM comparison.","marker":"(Hubert & Arabie, 1985)"}],"fun_headline_variants":["Dual k-means co-clusters speeches and terms","Cosine co-clustering splits inaugurals into two eras","Lincoln's 1865 speech joins the modern cluster","Joint word-document clustering reveals historical split","SDKM clusters speeches and vocabulary together"],"cache_read_input_tokens":21120,"weakest_assumption_plain":"The paper assumes that comparing documents and terms by raw inner products is the same as comparing them by angle, even though the TF-IDF matrix is not normalized before the assignment steps, and that the row-normalized centroid update really maximizes the stated objective.","fun_headline_variants_meta":{"raw":{"variants":["Dual k-means co-clusters speeches and terms","Cosine co-clustering splits inaugurals into two eras","Lincoln's 1865 speech joins the modern cluster","Joint word-document clustering reveals historical split","SDKM clusters speeches and vocabulary together"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000333,"raw_usage":{"total_tokens":1922,"prompt_tokens":1088,"completion_tokens":834,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":761}},"tokens_in":704,"tokens_out":834,"duration_ms":8152,"temperature":1.0,"reasoning_tokens":761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:29:33.212767+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SDKM twice on the same TF-IDF matrix, once exactly as Algorithm 1 prints it and once with every row and every column normalized to unit length before the iterations. If the two runs produce different cluster assignments or objective trajectories, the inner-product steps in Algorithm 1 are not equivalent to cosine similarity, and the paper's central premise fails.","supporting_citations":[{"cited_title":"( year 2001 )","cited_arxiv_id":null,"evidence_quote":"Introduces double k-means, the $X = U Y V' + E$ block model and alternating update scheme that SDKM inherits."},{"cited_title":", & author Vichi, M","cited_arxiv_id":null,"evidence_quote":"Supplies the pseudo-F index used to choose the numbers of row and column clusters."},{"cited_title":", & author Arabie, P","cited_arxiv_id":null,"evidence_quote":"Provides the Adjusted Rand Index used to evaluate partition recovery in the simulations and the DKM comparison."}],"review_version":1}