{"id":"60f939c3-064e-4106-8d63-f20b6195ad78","arxiv_id":"2507.12029","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A matrix-factorization method that learns view weights from labeled known classes and transfers them to cluster novel classes in multi-view data.","lead":"This paper proposes a new method for discovering novel classes when data comes from multiple views, such as gene expression and imaging in medicine. It claims the first approach to this setting and reports accuracy gains over existing clustering methods on eight datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central effectiveness claim is not yet supported because λ1 and λ2 appear to be tuned using novel-class test labels, so Table 2 may overstate IICMVNCD; the reader's view-weight transfer concern is real but secondary.","rationale":"I examined whether the reader's weakest assumption, transfer of view weights from known to novel classes, is the load-bearing problem. It is not: in the actual optimization, α is updated in Eq. (13) using r_v = ||Xv - WvAvY||_F over the concatenated known and novel feature matrix, so α is influenced by novel samples from the start. The paper's prose in Sec. 3.3 and the abstract overstates a 'transfer' that is really simultaneous joint optimization; this is a narrative defect, not the main threat to the central claim. The main threat is that the experimental comparison may be invalid. Section 5.5's statement that tuning hyperparameters using known-class performance is future work is direct evidence that the authors selected λ1 and λ2 using the test (novel-class) labels, or at least did not follow a known-class validation protocol. Because the headline result is 'consistently outperforms' across eight datasets, and the margins over the strongest baselines are often small (e.g., Wiki ACC 65.42 vs 64.31; CCV ACC 34.20 vs 33.52; KIPAN ACC 92.51 vs 91.73), tuning two hyperparameters on the evaluation labels could easily explain these differences. The concrete test is straightforward and would settle the concern: re-evaluate with validation-based hyperparameter selection and multiple splits. I therefore keep the reader's CONDITIONAL verdict unchanged; the paper is promising but the central empirical claim needs verification under a non-leaking protocol.","tokens_in":14534,"tokens_out":8711,"duration_ms":101544,"concrete_test":"Rerun Table 2 with a leakage-free protocol: select λ1 and λ2 using only known-class information, e.g., maximize clustering accuracy or reconstruction loss on a held-out subset of known classes, then evaluate on novel classes. Repeat over multiple known/novel class splits and report mean±std. If the average gains over CKD, OPMC, and other baselines shrink or reverse, the central effectiveness claim fails. Also report the selected λ values per dataset to assess sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that IICMVNCD clusters novel classes in multi-view data better than MVC and NCD baselines, and Table 2 is the evidence. Section 5.1.2 says λ1 and λ2 are tuned from 10^[0,1,...,5], but no validation split or known-class-based selection is described. Section 5.5 explicitly states 'we will explore how to tune them based on the results of known classes in future research,' which is an admission that the reported numbers were not produced under the protocol required for NCD, where novel-class labels are the test labels. With eight datasets and six candidate values per λ, selecting the combination that maximizes ACC/NMI/Purity on novel classes can materially inflate results and makes the 'consistently outperforms' claim unverifiable. The reader's transfer concern is secondary: Eq. (3) updates α in Eq. (13) from reconstruction error over all data, including novel samples, so the statement that view weights are learned from known classes and then applied to novel classes is not literally what the optimization does; the actual transfer happens through the shared basis and the disjointness penalty. The decisive issue for the central claim is the evaluation protocol.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes IICMVNCD, a matrix-factorization-based method for novel class discovery in multi-view data. The model decomposes each view into a shared orthonormal basis and a factor matrix over known and novel samples, generates one-hot labels via weighted fusion, and uses a known-class supervision term plus a novel-known separation penalty. An alternating optimization with closed-form updates is derived, and experiments on eight datasets compare with multi-view clustering and NCD baselines.","tokens_in":14807,"tokens_out":6252,"duration_ms":74191,"significance":"If the reported gains hold, the paper fills a genuine gap: NCD has been studied almost exclusively in single-view settings, and the proposed closed-form updates and linear-in-n complexity are attractive. The derivations for W_v, A_v, and Y are mostly sound, and the authors provide a convergence argument, ablation study, and parameter-sensitivity analysis. However, the evaluation protocol and an inconsistency in the alpha update currently prevent me from accepting the central empirical claim.","major_comments":[{"comment":"The hyperparameters lambda1 and lambda2 are tuned from a grid (10^[0,1,...,5]) with no validation split described, and Section 5.5 states that tuning them on known classes is future work. Because the novel-class labels are the test labels in NCD, selecting the hyperparameters on the test set means the reported ACC/NMI/Purity in Table 2 may be optimistically biased and the 'consistently outperforms' claim is not verifiable. Please provide a validation protocol, such as a held-out split of known classes or a known-class proxy, and report results under that protocol.","section":"5.1.2 and 5.5"},{"comment":"There is a mismatch between the objective in Eq. (3), which contains alpha_v/2 times the reconstruction error, and the alpha subproblem in Eq. (13), which minimizes sum_v alpha_v^2 r_v^2. The closed-form update in Eq. (15) solves the quadratic surrogate, not the stated objective, so the monotone-decrease argument in Eq. (17) does not follow for the original J. Please align the objective and the subproblem, or justify the quadratic surrogate as an intentional modification.","section":"3.4 and 4.4"},{"comment":"The paper repeatedly states that view weights are learned from known classes and then transferred to novel classes, but the update in Eq. (13) computes r_v^2 = ||X_v - W_v A_v Y||_F^2 over the full X_v = [X_l^v, X_u^v], including novel samples. Thus alpha is influenced directly by novel-class reconstruction errors, and the actual transfer happens through the shared basis and label matrix rather than through the weights themselves. Please correct the narrative and, ideally, test the transfer assumption by comparing alpha learned on known classes only versus the current global update.","section":"3.3 and 4.4"},{"comment":"The ablation claim that 'removing any one of the modules results in a decline' is not supported by the table: on STL10 the 'w/o alpha' variant already reaches 99.02 ACC, identical to the full method, and on Wiki and YTB10 the differences are very small (64.35 vs. 65.42 and 94.32 vs. 94.55). Moreover, on uci-digit 'w/o alpha' (95.20) is nearly as good as the full model (95.30). Please temper the claim or provide repeated-trial results with error bars.","section":"5.4, Table 3"}],"minor_comments":[{"comment":"The bound sum_{g_i in G_l} sum_{y_j in Y_u} ||g_i - y_j||_F^2 <= n_l n_u sqrt(2) is incorrect; each squared distance equals 2, so the sum is 2 n_l n_u. The lower-bound constant in Eq. (19) should be updated accordingly.","section":"4.5.1, Eq. (18)"},{"comment":"The acronym is written as 'IICMVNVD' in the introduction, but 'IICMVNCD' is used elsewhere.","section":"1"},{"comment":"The sentence 'These findings highlight the importance of developing NCD techniques that can fully leverage multi-view information...' is duplicated verbatim at the end of the section.","section":"5.2"},{"comment":"The adaptation of NCD baselines by replacing their backbone with separate MLPs and concatenating features may disadvantage methods designed for image augmentations; please specify whether any image-specific preprocessing was used and discuss this limitation.","section":"5.1.2"},{"comment":"The notation '10.^[0, 1, ..., 5]' and the axis labels in Figure 3 are garbled; please clarify the grid and the plotted metric.","section":"5.5"}],"recommendation":"major_revision","confidential_remarks":"The 'first attempt at multi-view NCD' claim should be verified against the broader literature, including recent multi-view clustering and generalized NCD works, since the authors' group has extensive related publications. The main review risk is the test-set hyperparameter selection, which should be fixed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is the first multi-view NCD formulation I know, and the optimization is mostly solid, but the empirical claims rest on hyperparameters tuned with the test labels, so Table 2 doesn't yet demonstrate what it claims.\n\nWhat's new: the problem setting — novel class discovery with multiple views — has not been addressed in the cited NCD or MVC literature, and the paper makes a reasonable case it matters (multi-omics diagnosis is a genuine example). The method itself is a coherent recombination of shared-basis matrix factorization and a view-weighting scheme, with closed-form alternating updates and a monotonic-decrease argument. That's a real contribution in a subfield that is moving fast.\n\nThe soft spots, in proportion. The load-bearing one is the evaluation protocol. λ1 and λ2 are tuned from a grid with no validation split, and Section 5.5 states that tuning via known classes is future work — i.e., the reported numbers were selected on the novel-class test labels. With six values per λ and eight datasets, that can materially inflate the 'consistently outperforms' claim. The stress-test note got this right.\n\nThe view-weight transfer story is also overstated. In the actual update, α is recomputed from the total reconstruction error over both known and novel data (Eq. 13), not from the known-class supervision loss. Transfer happens through the shared basis and the disjointness penalty, but not the literal known-to-novel weight transfer described in Sec. 3.3. That is a representational flaw in the paper, not necessarily a fatal one in the method.\n\nMinor but real: the ablation is mixed — on STL10 and YTB10 the α module adds little or nothing — and the bound in Eq. (18) is off by a constant (max squared distance between one-hot vectors is 2, not √2), so the convergence proof has a small error even though the monotonic decrease + lower bound argument still works. No error bars, and a single split with the first half/second half class ordering.\n\nWho it's for: NCD and MVC researchers, especially those doing multi-omics work. It deserves a serious referee, but only with major revision: redo the tuning protocol, report variance over multiple splits, and correct the transfer narrative. If the numbers hold up under a valid protocol, it's a solid paper.","headline":"First multi-view NCD formulation with a coherent optimization, but the headline accuracy claims are undercut by test-set hyperparameter tuning and an overstated view-weight transfer story.","tokens_in":15307,"tokens_out":3531,"would_cite":true,"duration_ms":41184,"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":"The paper's central claim is that novel class discovery can be extended to multi-view data by transferring view weights learned from known classes to novel classes, and that this gives better clustering than existing multi-view clustering…","keywords":["novel class discovery","multi-view clustering","matrix factorization","view weights","pseudo-label-free","multi-omics","clustering","label transfer"],"falsifier":"Construct a two-view dataset where view 1 cleanly separates the known classes but view 2 alone separates the novel classes, then run IICMVNCD: if view 1 keeps its high weight on the novel partition and accuracy drops below an equal-weight or oracle-weight baseline, the transfer assumption is refuted.","tokens_in":14387,"feed_emoji":"🧩","tokens_out":6893,"duration_ms":71740,"temperature":0.7,"pith_summary":"Novel class discovery asks an algorithm to cluster unlabeled data whose categories are disjoint from a small labeled set, and until now this task has been studied almost exclusively on single-view data. This paper argues that the natural next step is multi-view novel class discovery, where each sample has several complementary views, and it proposes IICMVNCD as the first method aimed at this setting. The central claim is that knowledge from the labeled classes can be transferred in two coordinated ways: within each view, a shared matrix-factorization basis captures the distributional overlap between known and novel classes, and across views, the importance weights learned from known-class supervision guide how the novel classes are fused and labeled. If this transfer works, multi-view NCD can be solved without pseudo-labels, which the paper identifies as a major source of instability in existing NCD methods. Experiments across eight multi-view datasets, including multi-omics data, support that claim with gains in clustering accuracy, NMI, and purity.","feed_headline":"View weights learned on known classes now guide novel-class clustering","feed_subtitle":"Matrix factorization fuses views while supervision from labeled classes sets weights, beating baselines on eight benchmarks.","key_machinery":"The carrying object is a per-view matrix factorization $X_v = W_v A_v Y$ with an orthonormal shared basis $W_v \\in \\mathbb{R}^{d_v \\times k}$, a centroid matrix $A_v$, and a one-hot predicted label matrix $Y = [Y_l, Y_u]$, together with a view-weight vector $\\alpha$ constrained to the simplex. The shared basis is what transfers distributional knowledge between known and novel classes; the weighted fusion $\\sum_v \\alpha_v \\|X_v - W_v A_v Y\\|_F^2$ is how inter-view correlation is expressed; and the supervision term $\\lambda_1 \\|Y_l - G_l\\|_F^2$ is what learns $\\alpha$ from labeled data before it is handed to the novel partition. Updates alternate in closed form: $W_v$ by SVD, $A_v$ by a linear solve, $Y$ by a discrete per-sample assignment, and $\\alpha$ by a Cauchy-Schwarz weighted inverse-residual formula.","core_discovery":"On its own terms, the paper establishes that novel classes in multi-view data can be clustered by borrowing two kinds of structure from known classes. At the intra-view level, features of labeled and unlabeled samples are concatenated and decomposed into a view-specific orthonormal basis matrix shared by both sets and a factor matrix that records pairwise sample relationships; the shared basis is what carries distributional consistency between known and novel classes. At the inter-view level, the factor matrices are fused into one-hot predicted label matrices with per-view weights, and those weights are adjusted so that the predicted labels of known classes match their ground truth. The learned weight vector is then applied directly to the novel-class reconstruction and label assignment, and a repulsion term keeps novel labels disjoint from known labels. The paper's headline empirical claim is that this scheme beats both multi-view clustering baselines and adapted single-view NCD methods on all eight benchmark datasets.","pith_inferences":["An untested but testable extension is to verify view-weight transfer directly, e.g., by measuring each view's clustering accuracy on known versus novel classes and checking whether the relative ranking matches the learned $\\alpha$.","The framework operates on raw feature matrices, so it could plausibly be composed with learned deep embeddings or graph views; the paper does not explore either.","The disjointness repulsion term suggests a natural stress test: datasets where known and novel classes are close in feature space, where that term would need to do most of the work.","Because the paper tunes $\\lambda_1$ and $\\lambda_2$ on labeled data, automated selection from known-class performance is a plausible practical follow-up, which the authors themselves flag."],"forward_implications":["Multi-view novel class discovery becomes a real setting rather than an implicit single-view assumption, so applications such as multi-omics disease discovery can treat each omics layer as a view.","Pseudo-label supervision can be dropped from the multi-view pipeline, removing a known source of instability in existing NCD methods.","The learned view weights are interpretable: they say which data view matters most for separating known classes, and the paper assumes that ranking carries over to novel classes.","Because each iteration is linear in the number of samples and convergence is guaranteed, the method is scalable to large datasets like the 38,654-sample video dataset used in evaluation.","On the eight reported benchmarks, the method achieves the best ACC, NMI, and Purity against both multi-view clustering and NCD baselines."],"supporting_citations":[{"why":"Formalizes novel class discovery with disjoint known and novel categories, the problem definition this paper adopts.","marker":"[11]"},{"why":"Represents the pseudo-label-based one-stage NCD line that this paper argues is unstable in multi-view data.","marker":"[12]"},{"why":"The IIC baseline that models inter-class and intra-class constraints, which this paper adapts to inter-view guidance.","marker":"[24]"},{"why":"The CKD knowledge-distillation baseline used to show that existing single-view NCD methods transfer poorly to multi-view data.","marker":"[10]"},{"why":"Supplies the joint non-negative matrix factorization formalism that the intra-view decomposition extends.","marker":"[25]"},{"why":"Provides the SVD closed-form update used to solve the W_v subproblem.","marker":"[51]"},{"why":"An auto-weighted multi-view clustering baseline whose view-weighting idea this paper supervises with known-class labels.","marker":"[43]"},{"why":"A one-pass matrix-factorization multi-view clustering baseline that the experiments must beat.","marker":"[27]"}],"fun_headline_variants":["Multi-view novel class discovery via shared bases and view weights","Intra- and inter-view cues guide novel class clustering in multi-view data","Novel classes in multi-view data discovered using known-class structure","Matrix factorization transfers view weights to novel-class clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole transfer rests on the premise that the relative reliability of views is the same for known and novel classes; if the views inform the two sets differently, the weights learned from labeled data will misdirect the unlabeled clustering.","fun_headline_variants_meta":{"raw":{"variants":["Multi-view novel class discovery via shared bases and view weights","Intra- and inter-view cues guide novel class clustering in multi-view data","Novel classes in multi-view data discovered using known-class structure","Matrix factorization transfers view weights to novel-class clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3024,"prompt_tokens":982,"completion_tokens":2042,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":1972}},"tokens_in":598,"tokens_out":2042,"duration_ms":17537,"temperature":1.0,"reasoning_tokens":1972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:55:53.389224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a two-view dataset where view 1 cleanly separates the known classes but view 2 alone separates the novel classes, then run IICMVNCD: if view 1 keeps its high weight on the novel partition and accuracy drops below an equal-weight or oracle-weight baseline, the transfer assumption is refuted.","supporting_citations":[{"cited_title":"Learning to discover novel visual categories via deep transfer cluster- ing","cited_arxiv_id":null,"evidence_quote":"Formalizes novel class discovery with disjoint known and novel categories, the problem definition this paper adopts."},{"cited_title":"Automatically discov- ering and learning new visual categories with ranking statis- tics","cited_arxiv_id":null,"evidence_quote":"Represents the pseudo-label-based one-stage NCD line that this paper argues is unstable in multi-view data."},{"cited_title":"Modeling inter-class and intra-class constraints in novel class discov- ery","cited_arxiv_id":null,"evidence_quote":"The IIC baseline that models inter-class and intra-class constraints, which this paper adapts to inter-view guidance."},{"cited_title":"Class- relation knowledge distillation for novel class discovery","cited_arxiv_id":null,"evidence_quote":"The CKD knowledge-distillation baseline used to show that existing single-view NCD methods transfer poorly to multi-view data."},{"cited_title":"Multi-view clustering via joint nonnegative matrix factorization","cited_arxiv_id":null,"evidence_quote":"Supplies the joint non-negative matrix factorization formalism that the intra-view decomposition extends."},{"cited_title":"Multi-view clustering via late fusion alignment maximization","cited_arxiv_id":null,"evidence_quote":"Provides the SVD closed-form update used to solve the W_v subproblem."},{"cited_title":"Auto- weighted multi-view clustering for large-scale data","cited_arxiv_id":null,"evidence_quote":"An auto-weighted multi-view clustering baseline whose view-weighting idea this paper supervises with known-class labels."},{"cited_title":"One-pass multi- view clustering for large-scale data","cited_arxiv_id":null,"evidence_quote":"A one-pass matrix-factorization multi-view clustering baseline that the experiments must beat."}],"review_version":1}