{"id":"59f72bad-3017-4140-bb26-6436222cd2b8","arxiv_id":"2607.05635","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"IFGRVFL-MV, fusing intuitionistic fuzzy weights, graph embedding and multiview learning into RVFL, reaches 81.06% average accuracy versus 79.5% for the closest prior multiview graph RVFL on eight UCI/KEEL sets.","lead":"The paper introduces IFGRVFL-MV, an RVFL variant that adds intuitionistic fuzzy sample weights, graph embedding, and multiview fusion for classification under noise. It reports higher average accuracy than four baselines on eight small UCI/KEEL datasets and may interest practitioners building classifiers for multi-source uncertain data.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Non-significant Wilcoxon vs GRVFL-MV plus tiny dataset suite leave the superiority claim under-supported.","rationale":"The Reader correctly isolates the weakest link: the non-significant Wilcoxon result against GRVFL-MV together with the very small collection of modest-sized data sets. The mathematical derivation of the closed-form solution (Eqs. 28–40) is transparent and the three ingredients are cleanly combined, so correctness of the algorithm itself is not in doubt. The load-bearing issue is purely empirical: without a statistically significant pairwise improvement over the nearest baseline, the claim of out-performance cannot be regarded as established. My concrete test simply enlarges the evidence base under the same protocol; a still-non-significant p-value would confirm that the present results are under-powered, while a significant p-value would remove the concern. Because the Reader already assigned CONDITIONAL for precisely this reason, no change of verdict is required.","tokens_in":13966,"tokens_out":511,"duration_ms":8755,"concrete_test":"Re-run the entire experimental suite on at least 20 additional UCI/KEEL (or OpenML) binary classification sets of comparable size, keep the same hyper-parameter grid and 70/30 protocol, recompute the Wilcoxon signed-rank p-value of IFGRVFL-MV versus GRVFL-MV; if that p-value remains > 0.05 the superiority claim is not statistically supported and the paper’s strongest claim must be softened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Abstract; Table I; §IV-C) is that IFGRVFL-MV is superior, with average accuracy 81.06 % and Friedman rank 1.19. That claim rests on the assumption that the eight UCI/KEEL sets (n = 102–961) plus grid-searched hyper-parameters and a single 70/30 split are enough to establish general out-performance. Yet the decisive pairwise Wilcoxon signed-rank test against the immediate predecessor GRVFL-MV yields p = 0.195 (Table II), which fails to reject the null at any conventional α. Absolute gains are also modest (1.56 points over GRVFL-MV). Consequently the statistical evidence does not yet underwrite the unqualified superiority asserted in the abstract and conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes IFGRVFL-MV, an extension of GRVFL-MV that adds intuitionistic fuzzy membership/non-membership scores (via class-center distances and neighbor heterogeneity) into a multiview RVFL objective that already includes graph-embedding regularizers and a cross-view error consistency term. The resulting constrained problem (Eq. 28) is solved in closed form after forming the Lagrangian (Eqs. 29–40), yielding a block linear system for the two view-specific output weights. Experiments on eight UCI/KEEL binary classification sets (n = 102–961) with 70/30 splits and grid-searched hyperparameters report average accuracy 81.06 % and Friedman rank 1.19, outperforming RVFL1/2, GE-IFWRVFL and GRVFL-MV; Friedman, Nemenyi, Wilcoxon and win-tie-loss statistics are supplied.","tokens_in":14246,"tokens_out":1078,"duration_ms":8387,"significance":"If the superiority claim holds, the work supplies a practical, closed-form multiview RVFL that jointly handles noise via intuitionistic fuzzy weights and geometry via graph embedding—useful for small-to-medium tabular problems where uncertainty and complementary views matter. The algebraic derivation is transparent and the model is a natural, incremental combination of three existing strands (IFS, GE, MVL). Strengths include an explicit closed-form solution and a full suite of non-parametric tests. The contribution remains incremental rather than foundational; its practical impact hinges on whether the modest gains generalize beyond the eight small benchmarks examined.","major_comments":[{"comment":"Abstract, §IV-C and Table II: the central claim that IFGRVFL-MV “outperforms existing models” is not statistically supported against the immediate predecessor. The Wilcoxon signed-rank p-value versus GRVFL-MV is 0.195 (non-significant at any conventional α); absolute average gain is only 1.56 points. With N = 8 the Friedman/Nemenyi results alone cannot underwrite the unqualified superiority asserted in the abstract and conclusion. Either enlarge the suite substantially, report confidence intervals / effect sizes, or temper the claim to “competitive / modest improvement.”","section":"Abstract; §IV-C; Table II"},{"comment":"§IV-A and Table I: all eight datasets are small (n ≤ 961) and only a single 70/30 split is used. Grid search over many free parameters (h, c1–c4, θ1, θ2, ρ, μ, η) on such limited data raises a clear risk of optimistic bias. Nested cross-validation or repeated random splits with reported variance are needed before the average-accuracy ranking can be treated as reliable.","section":"§IV-A; Table I"},{"comment":"§III, Eq. (28) and surrounding text: the construction of the two views (P and Q) is never specified for the UCI/KEEL tables that are originally single-view. Without an explicit, reproducible view-generation protocol (feature split, random projection, etc.) the multiview component cannot be independently verified or compared fairly with single-view baselines.","section":"§III; Eq. (28)"}],"minor_comments":[{"comment":"Table I caption says “AUC values” while the columns report accuracy; correct the caption.","section":"Table I"},{"comment":"Eq. (41) writes K(z,y) but the right-hand side uses ∥z−x∥; fix the dummy variables.","section":"§IV-A, Eq. (41)"},{"comment":"Notation for the GE matrices switches between A_w, A_w1/A_w2 and A^P_w/A^Q_w without a clear definition of how the fuzzy scores enter the graph Laplacian; a short clarifying sentence would help.","section":"§III"},{"comment":"Several self-citations appear as arXiv preprints; if journal versions exist they should be preferred.","section":"References"},{"comment":"Typographical slips: “utilizes” → “utilize” (contributions list), “dimen-sions”, “P atterns×F eatures” in Table I header.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent but incremental combination of three already-published RVFL extensions (IFS, GE, MVL). The non-significant Wilcoxon against the closest baseline and the tiny dataset suite are the decisive weaknesses; if the authors cannot enlarge the empirical base or soften the claim, the paper is better suited to a specialized workshop or a lower-tier venue than a top ML journal. No integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a straightforward engineering paper that sticks three existing pieces—intuitionistic fuzzy sample weights, graph embedding, and multiview fusion—into one RVFL objective (Eq. 28) and derives the closed-form solution (Eq. 40). That joint model and the accompanying numbers on eight UCI/KEEL sets are new; each separate technique already appears in the cited IFRVFL, GE-IFWRVFL and GRVFL-MV papers.\n\nWhat they do well is the algebra. The Lagrangian and the block-matrix solution look consistent, the notation is readable, and they report Friedman ranks, Nemenyi CD, Wilcoxon, and win-tie-loss tables rather than just raw averages. Average accuracy moves from 79.5 % (GRVFL-MV) to 81.06 %, and the Friedman rank is best. That is a usable incremental result inside the randomized-network literature.\n\nThe soft spot is exactly where the stress-test note points: the decisive Wilcoxon against the immediate predecessor GRVFL-MV is p = 0.195. With only eight small sets (n = 102–961), a single 70/30 split, and a large grid of free parameters (hidden sizes, six regularizers, kernel width, fuzzy threshold), the abstract’s unqualified “outperforms existing models” claim is not yet underwritten. Absolute gains are modest. No code is released. These are real but proportionate limitations; they do not break the derivation or invent circularity.\n\nThe paper is for people already working on RVFL variants who want a ready-to-implement multi-view fuzzy-graph version. A serious referee should see it; the math is clean enough and the empirical package is standard for the subfield. I would not cite it myself unless I am writing a survey of RVFL extensions, but I would not desk-reject it either.","headline":"Clean closed-form fusion of three known RVFL ingredients; the superiority claim over the nearest baseline is not statistically supported.","tokens_in":14824,"tokens_out":472,"would_cite":false,"duration_ms":4293,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Combining intuitionistic fuzzy weights, graph embedding and multiview learning inside RVFL yields higher classification accuracy on noisy multi-feature data.","keywords":["Random Vector Functional Link","intuitionistic fuzzy sets","graph embedding","multiview learning","classification","uncertainty handling"],"falsifier":"Train IFGRVFL-MV and GRVFL-MV on a larger, noisier multi-view collection (for example image-plus-text or multi-omics data with several thousand samples) under the same hyper-parameter protocol; if the accuracy gap disappears or reverses, the claimed advantage fails.","tokens_in":14882,"feed_emoji":"🔗","tokens_out":578,"duration_ms":5314,"temperature":0.7,"pith_summary":"Random Vector Functional Link networks train quickly by fixing random hidden weights and solving for the output weights in closed form, yet they struggle with noisy samples and with data that arrive as several complementary feature sets. This paper claims that three ingredients together fix both problems: intuitionistic fuzzy scores that down-weight uncertain points, graph-embedding regularizers that keep local geometry, and multiview learning that fuses two feature representations. The resulting IFGRVFL-MV model is trained by a single linear system whose solution is the pair of output weight matrices. On eight UCI and KEEL classification sets the model records the highest average accuracy (81.06 %) and the best Friedman rank among five RVFL variants, indicating that the combined treatment of uncertainty, geometry and multiple views improves generalization.","feed_headline":"Fuzzy multiview RVFL tops accuracy on noisy classification sets","feed_subtitle":"Closed-form model that weights uncertain points and keeps geometry beats four RVFL baselines.","key_machinery":"IFGRVFL-MV objective: a joint quadratic program over two view-specific output weights that penalizes fuzzy-weighted residuals, enforces graph-embedding regularizers and couples the views by a cross-term on the residuals; the solution is obtained by inverting one block matrix of size equal to the concatenated feature dimension.","core_discovery":"The authors show that adding intuitionistic-fuzzy sample weights and graph-embedding penalties to a multiview RVFL produces a closed-form classifier that outperforms plain RVFL, graph-embedded fuzzy RVFL and multiview graph RVFL on standard classification benchmarks, reaching 81.06 % average accuracy and Friedman rank 1.19.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Intuitionistic fuzzy multiview RVFL lifts accuracy on UCI sets","Graph-embedded IF-RVFL with multiview edges out baselines","Fuzzy weights and geometry push multiview RVFL to rank 1","Closed-form IFGRVFL-MV tops fuzzy and multiview RVFL variants","Multiview RVFL gains from intuitionistic fuzzy graph embedding"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that the combined model is generally superior rests on eight small public data sets and a non-significant Wilcoxon comparison against its immediate multiview predecessor.","fun_headline_variants_meta":{"raw":{"variants":["Intuitionistic fuzzy multiview RVFL lifts accuracy on UCI sets","Graph-embedded IF-RVFL with multiview edges out baselines","Fuzzy weights and geometry push multiview RVFL to rank 1","Closed-form IFGRVFL-MV tops fuzzy and multiview RVFL variants","Multiview RVFL gains from intuitionistic fuzzy graph embedding"]},"model":"grok-4.5","effort":"low","cost_usd":0.003774,"raw_usage":{"total_tokens":1137,"prompt_tokens":714,"num_sources_used":0,"completion_tokens":82,"cost_in_usd_ticks":37740000,"prompt_tokens_details":{"text_tokens":714,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":341,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":714,"tokens_out":82,"duration_ms":3328,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T04:33:45.135444+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train IFGRVFL-MV and GRVFL-MV on a larger, noisier multi-view collection (for example image-plus-text or multi-omics data with several thousand samples) under the same hyper-parameter protocol; if the accuracy gap disappears or reverses, the claimed advantage fails.","supporting_citations":[],"review_version":1}