{"id":"0c98eb1a-93a9-42a2-bd69-e1dc8e3a7f84","arxiv_id":"2509.00546","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper merges numerical and textual similarities in spectral clustering with a weight tuned on approved/rejected loan labels, yet the resulting 'recruitment implies lower risk' finding is contaminated by that same label information.","lead":"A method that blends financial ratios and loan audit texts for clustering SME borrowers is proposed, but its headline improvements and risk insights are built on a circular use of the bank's own approval labels. General risk professionals should read it as an application note, not as evidence that 'recruitment' predicts low default risk.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '30% lower default risk' claim is circular or undefined: the only risk-relevant labels (approved/rejected) are the same labels used to fit λ in Eq. (7), and no separate default outcome is defined.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the approval/rejection labels used to fit λ in Eq. (7) are reused to interpret cluster risk, making the headline 'recruitment predicts lower default risk' internally circular. My stress test sharpens this by noting that no separate default outcome is ever defined in the manuscript; the only risk-relevant labels are the same 1,428/438 labels used for constrained similarity optimization. I agree with the reader that this contaminates the central application claim. Other issues (unfair k comparison, Eq. (10) not actually run, corrupted equations, omitted prior baseline) also matter, but the risk-label circularity is the most load-bearing because it directly invalidates the paper's practical contribution. Since the reader's verdict is REJECT and this concern reinforces that verdict, I recommend no change to the verdict.","tokens_in":13379,"tokens_out":7251,"duration_ms":88794,"concrete_test":"Ask the authors to state exactly how '30% lower default risk' was computed in Section 5, including the outcome variable, formula, and sample. If no default/performance outcome exists, the claim is unevaluable and should be retracted. If a proxy exists, test independence of the discovery: split the 1,866 samples into training (e.g., 60%) and test (40%) partitions; fit λ in Eq. (7) using only the training approval/rejection labels; freeze W; cluster; and measure Cluster 3's social-recruitment incidence and risk difference in the test partition. Repeat over random splits. If the 30% effect is not consistently reproduced out-of-sample, it is an artifact of fitting λ to the approval labels rather than an independent cluster property.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 defines the 1,428 approved SMEs as the must-link set and the 438 rejected SMEs as the cannot-link set, and these labels are used to optimize λ in Eq. (7), yielding λ=0.65 for the combined similarity in Eq. (6). The paper's headline application insight (Abstract; Section 5) is that Cluster 3 is a 'low-risk' cluster, with 51% social-recruitment incidence and 30% lower default risk. For that insight to be an independent empirical discovery, the risk status of Cluster 3 must come from an outcome variable that was not used to construct the similarity matrix W. No such outcome is defined: the dataset is described only as 1,428 approved and 438 rejected loans. If 'low risk' is just another name for 'approved', then the same information that selected λ and shaped the spectral embedding is being reused to interpret the clusters, so the recruitment-risk association is a byproduct of the constraint optimization rather than a genuine finding. If '30% lower default risk' is computed from some other variable, the paper never reports what that variable is or how it was calculated. Either way, the credit-monitoring application claim is unsupported. This is more load-bearing than the silhouette-comparison issue because the paper's stated practical contribution depends on this cluster-level risk interpretation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Advanced Spectral Clustering (ASC), a method that fuses Mahalanobis-distance-based financial similarity with TF-IDF-based textual similarity through a weight λ, and selects the number of clusters via an 'eigenvalue-silhouette' optimization. The method is applied to a loan-audit dataset of 1,866 Chinese SMEs (1,428 approved, 438 rejected). The authors claim that ASC achieves an 18% higher Silhouette score than single-type baselines, that it is robust across k-means, k-medians, and k-medoids, and that the resulting clusters support credit-risk monitoring—specifically, that Cluster 3 contains 51% 'social recruitment' mentions and is associated with 30% lower default risk. The paper also compares ASC against several recent multi-view spectral clustering baselines using Silhouette, Calinski-Harabasz, and Davies-Bouldin indices.","tokens_in":13667,"tokens_out":4657,"duration_ms":56046,"significance":"If the claims were supported, ASC would offer a practically useful way to combine structured financial ratios with unstructured textual survey records for unsupervised credit monitoring. The paper has some strengths: it addresses a genuine problem (heterogeneous data fusion for spectral clustering), it provides a public data link, it reports comparisons with five state-of-the-art methods, and it includes robustness checks across three clustering algorithms. However, the central risk-related finding is undermined by the way the approved/rejected labels are used both to fit λ and then to interpret clusters as low-risk. In addition, the 18% Silhouette improvement is not an apples-to-apples comparison, and the key equations defining λ and k selection are too garbled to be reproducible. The methodological idea may have potential, but the evidence presented in this manuscript does not currently support the headline claims.","major_comments":[{"comment":"The risk interpretation is circular. Section 4.2 defines the 1,428 approved SMEs as the must-link set and the 438 rejected SMEs as the cannot-link set, and these labels are used to fit λ in Eq. (7). Section 5 then interprets Cluster 3 as having a '30% reduction in default risk' and the Abstract calls it 'low-risk.' No default-outcome variable is ever defined or computed; the only risk-related labels in the paper are the same approved/rejected labels that entered the λ optimization. The 'social recruitment implies lower default risk' claim is therefore not an independent empirical discovery. A concrete test: report the actual default outcome and recompute cluster risk without using any label that contributed to Eq. (7); if no such outcome exists, the claim should be removed or reframed as descriptive only.","section":"Section 4.2 and Section 5 (Eq. 7)"},{"comment":"The '18% higher Silhouette' claim compares ASC with k=3 against single-type baselines with k=4 (numerical) and k=11 (textual). Silhouette scores generally depend on the number of clusters, so this is not a controlled comparison. The paper itself states in Section 4.2.1 that 'it is not feasible to compare the baseline models with the advanced spectral clustering model using internal evaluation metrics such as intra-class distance and inter-class distance' because the optimal k differs. The 18% claim in the Abstract is incompatible with that caveat. Report Silhouette at the same k across methods, or provide a statistical comparison that justifies the claim.","section":"Section 4.2.1 and Abstract"},{"comment":"The two core optimization equations are not well-defined as printed. Eq. (7) contains an incomplete maximization expression, ambiguous constraints, and no explanation of how the objective is solved (beyond 'intervals of 0.05') or how Table 3 is derived from it. Eq. (10) is garbled: '||min max( ) max{ , } ii i k ii ACe AC− Δ+' is not a valid mathematical expression, and A_i and C_i are not defined precisely. Since λ selection and k selection are the paper's main claimed contributions, these issues make the method irreproducible. The authors should rewrite both equations and provide a step-by-step optimization procedure.","section":"Section 3.1, Eq. (7), and Section 3.2, Eq. (10)"},{"comment":"There is a data mismatch regarding textual data. Section 4.1 says the dataset contains 1,428 approved and 438 rejected SMEs, but Section 4.2 says the survey texts of the 1,428 SMEs serve as textual data and that Eq. (5) is applied to the 1,428 approved SMEs only. The 438 rejected SMEs are used as the cannot-link set in Eq. (7), but it is not explained whether their textual data exist, whether their similarity is computed only from financial variables, or how missing text is handled. This affects the construction of the fused similarity matrix and the λ optimization, so it must be clarified.","section":"Section 4.1 and Section 4.2"},{"comment":"The robustness claim in the Abstract, 'ΔSilhouette Coefficient < 0.02', is contradicted by Table 6, which reports Silhouette coefficients of 0.73, 0.75, and 0.77 for k-medians, k-means, and k-medoids respectively; the range is 0.04. The ΔIntra/Inter < 0.13 claim also relies on a single rounded comparison (0.7827 − 0.6528 = 0.1299), and no statistical significance test is provided for the claim that the differences are 'not statistically significant.' The robustness claim should be restated accurately or supported with proper tests.","section":"Abstract and Table 6"}],"minor_comments":[{"comment":"The table heading contains a typo: 'Tabel 1' should be 'Table 1.'","section":"Section 4.1"},{"comment":"'Consine measure' should be 'Cosine measure.'","section":"Section 3.1, Eq. (5)"},{"comment":"The word-frequency ratio formula is garbled: 'Number of the entity in the category / The total number of the entity' is not a well-defined mathematical expression. Please rewrite it with explicit summation indices.","section":"Section 4.1, Eq. (11)"},{"comment":"Figure 7 is referenced but no numerical values are reported in the text; the claim that ASC 'significantly outperforms' baselines on CHC and DBI is not supported by statistical tests or effect sizes.","section":"Section 4.3, Figure 7"},{"comment":"The final discussion paragraph about grid-based clustering is disconnected from the experiments; no grid-based method is tested. Either remove it or relate it to the reported results.","section":"Section 5"}],"recommendation":"reject","confidential_remarks":"The label-leakage issue is the central problem: the approved/rejected labels are used to fit λ and then reused to interpret Cluster 3 as 'low-risk' with a 30% default-risk reduction, despite no default outcome being defined. This cannot be fixed by rewriting alone; it requires either an external outcome variable or a substantial reframing of the contribution. Combined with the garbled core equations and the inconsistent Silhouette robustness claim, the manuscript in its current form does not meet the standard for publication. If the authors can obtain a genuine default/performance outcome and redo the analysis, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper combines two standard similarity measures—Mahalanobis on five financial ratios and TF-IDF cosine on audit text—via a scalar weight, and tunes that weight using approved/rejected labels as must-link/cannot-link constraints. The practical story is that clustering finds a 'social recruitment' cluster with lower default risk. The story doesn't hold up.\n\nThe dataset is real and the problem is worth working on. The robustness checks across k-means/k-medians/k-medoids are fine, and the paper is honest about its text-similarity limitations. But the main empirical claims are problematic.\n\nFirst, the circularity: λ is fit to separate the 1,428 approved from the 438 rejected. Then Cluster 3 is called 'low-risk' and the recruitment association is presented as a discovery. There is no separate default outcome defined anywhere, so '30% lower default risk' either comes from those same approval labels (circular) or from an unidentified variable (unsupported). The stress-test is right, and it's the load-bearing issue.\n\nSecond, the 18% Silhouette improvement compares ASC at k=3 with baselines at k=4 (numerical) and k=11 (text). That's not an apples-to-apples comparison, and the paper admits the differing k makes internal metrics non-comparable, then uses Silhouette anyway.\n\nThird, the novel k-selection in Eq. (10) is garbled and not actually used; the experiments rely on eigen-gap plus Silhouette. Eqs. (7) and (10) are barely readable, which is a reproducibility problem.\n\nFourth, the paper omits the authors' own prior fuzzy clustering on the same data (Han et al. 2023) as a baseline. That's a conspicuous omission.\n\nThere's also a data ambiguity: the paper says textual data are for the 1,428 approved SMEs, but the cannot-link set needs text for rejected firms too. That needs clarification.\n\nAll that said, this is not a crank paper. The authors are engaging a real problem and the core idea—using constrained labels to fuse modalities—has some merit, though it's not new. With a proper default outcome, controlled baseline comparisons, and cleaned-up equations, this could become a modest but useful applied paper. As it is, I would not accept it.\n\nFor peer review: send it out. It deserves referee time because the flaws are identifiable and fixable, and the domain is relevant. Just expect major revision or a reject after review.\n\nI wouldn't cite it in its current form. If you work on credit analytics, it's worth a quick skim as a cautionary example.","headline":"Simple weighted-similarity method on a real dataset, but the performance claims are misleading and the main risk insight is circular.","tokens_in":14190,"tokens_out":3916,"would_cite":false,"duration_ms":43528,"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 proposes an unsupervised spectral method that fuses financial ratios with loan-audit text, and reports that the fused clusters separate SME risk profiles better than either data type alone — including a recruitment-heavy cluster w","keywords":["advanced spectral clustering","heterogeneous data","credit risk monitoring","SME credit","Mahalanobis distance","TF-IDF cosine similarity","Silhouette coefficient","group portrait"],"falsifier":"Run ASC with λ chosen without access to approval labels — for example, by maximizing Silhouette alone on a training subset — and test whether Cluster 3 still shows ~51% 'social recruitment' and 30% lower default risk out-of-sample. A direct check would be a supervised model with the five financial ratios plus a social-recruitment text indicator: if the indicator adds no predictive value, the recruitment signal is not an independent discovery.","tokens_in":13239,"feed_emoji":"📊","tokens_out":10320,"duration_ms":104610,"temperature":0.7,"pith_summary":"The paper sets out to show that credit monitoring of small and medium-sized enterprises improves when loan data are treated as heterogeneous — combining numerical financial ratios with the free-text notes auditors write during quarterly inspections. It proposes Advanced Spectral Clustering (ASC), which builds one similarity matrix as a weighted sum of Mahalanobis (correlation-adjusted) distance on five financial ratios and normalized TF-IDF (term-frequency/inverse-document-frequency) cosine similarity on a 236-word audit-text lexicon. The weight is tuned so that approved and rejected loans separate cleanly, and the number of clusters is chosen by balancing eigenvalue gaps against a Silhouette-based criterion. On 1,428 SMEs, the method reports an 18% higher Silhouette score than single-type-data baselines, stable behavior across k-means, k-medians, and k-medoids, and a cluster rich in 'social recruitment' language whose members have 30% lower default risk.","feed_headline":"Fusing loan numbers and audit text beats single-type clustering by 18%","feed_subtitle":"A new spectral method finds a recruitment-focused SME cluster with 30% lower default risk, a cheap signal for banks.","key_machinery":"The load-bearing object is the fused similarity matrix W with entries λ·Sim_financial(x_i,x_j) + (1−λ)·Sim_textual(t_i,t_j). λ is not guessed: it solves the constrained optimization in Eq. (7), using approved loans as a must-link set and rejected loans as a cannot-link set. The second mechanism is the k-selection score in Eq. (10), which mixes eigenvalue gaps with an intra-cluster/inter-cluster ratio; this replaces the usual 'look for the elbow' heuristic for choosing how many eigenvectors to keep. These eigenvectors of the normalized graph Laplacian become the coordinates on which k-means (or k-medians/k-medoids) runs, so the final clusters live in a spectral embedding rather than in the ra","core_discovery":"ASC's central claim: fusing Mahalanobis financial similarity with TF-IDF textual similarity, and choosing both the fusion weight and the eigenvector count by optimization, makes spectral clustering yield useful SME credit profiles. The weight λ is fit so approved loans (must-link) resemble each other and rejected loans (cannot-link) stay separate (Eq. 7); the cluster count k balances eigenvalue gaps against cluster separation (Eq. 10), giving k=3. The clusters are interpretable — manufacturers, distress-signal firms, and a recruitment-dominated group — and the recruitment group is reported to have 30% lower default risk. This is taken as evidence that an unsupervised pipeline can convert aud","pith_inferences":["The approval/rejection labels used to fit λ may leak into the clusters, so the Cluster 3 default-risk finding is not fully independent until tested out-of-sample or with an unsupervised λ.","The 'social recruitment' association is a word-frequency correlation, not a controlled risk factor; a supervised test against the five financial ratios would show whether it adds predictive power.","The method's pairwise similarity and spectral decomposition scale quadratically/cubically, so applying it to large lender portfolios would require the community-level approximation the paper mentions as future work.","A natural testable extension is to replace the bag-of-words TF-IDF text similarity with semantic embeddings; the paper itself notes current text handling misses synonymous phrases, so the 51% recruitment signal may be diluted by poor synonym recognition."],"forward_implications":["If ASC's 18% Silhouette gain holds, banks can profile SME credit risk from routinely collected audit texts plus financial ratios rather than choosing one modality.","The k-selection rule removes the manual eigen-gap inspection, so the method can be applied by analysts who are not spectral-clustering specialists.","Cluster 3's recruitment signal, if it replicates, gives a low-cost screen: firms actively recruiting look like safer borrowers, lowering the need for on-site inspections.","Stability across k-means, k-medians, and k-medoids means the final grouping does not hinge on the last clustering step, supporting use in operational workflows."],"supporting_citations":[{"why":"Supplies the SME loan-audit text dataset and the entity/lexicon extraction that feeds the textual similarity.","marker":"Han et al. (2023)"},{"why":"Motivates using non-financial information in SME credit evaluation, the premise for adding textual data.","marker":"(Jiang et al., 2023)"},{"why":"Provides the degree-matrix and Laplacian formulation ASC's spectral step builds on.","marker":"Chifu et al. (2015)"},{"why":"Shows spectral clustering applied to text with cosine similarity, supporting the textual-similarity design.","marker":"Roy & Basu (2022)"},{"why":"Defines the LD-SSC baseline ASC must beat in the comparison experiments.","marker":"(Yan, Shen, & Wang, 2014)"},{"why":"Defines the RMSC multi-view spectral baseline used in the comparison.","marker":"(Xia et al., 2014)"},{"why":"Defines the PSC pairwise-consistency baseline used in the comparison.","marker":"(Lu, Yan, & Lin, 2016)"},{"why":"Defines the U-SPEC ultra-scalable spectral baseline used in the comparison.","marker":"(Huang et al., 2020)"},{"why":"Defines the FUMC fast unpaired multi-view clustering baseline used in the comparison.","marker":"(Li et al., 2024)"}],"fun_headline_variants":["Fusing numbers and text lifts credit clustering 18%","Text + numbers spectral clustering finds low-risk SMEs with 30% less default","Spectral clustering on mixed data spots recruitment niche with lower default risk","Blending financial and textual features improves credit risk clusters by 18%","New spectral method merges data types, finds recruitment SMEs with 30% lower risk"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the approved-versus-rejected loan records are an unbiased risk signal, so tuning the fusion weight to separate them does not simply bake that signal into the clusters and then rediscover it in Cluster 3.","fun_headline_variants_meta":{"raw":{"variants":["Fusing numbers and text lifts credit clustering 18%","Text + numbers spectral clustering finds low-risk SMEs with 30% less default","Spectral clustering on mixed data spots recruitment niche with lower default risk","Blending financial and textual features improves credit risk clusters by 18%","New spectral method merges data types, finds recruitment SMEs with 30% lower risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000754,"raw_usage":{"total_tokens":3185,"prompt_tokens":733,"completion_tokens":2452,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":2356}},"tokens_in":477,"tokens_out":2452,"duration_ms":18346,"temperature":1.0,"reasoning_tokens":2356,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:28:29.256362+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ASC with λ chosen without access to approval labels — for example, by maximizing Silhouette alone on a training subset — and test whether Cluster 3 still shows ~51% 'social recruitment' and 30% lower default risk out-of-sample. A direct check would be a supervised model with the five financial ratios plus a social-recruitment text indicator: if the indicator adds no predictive value, the recruitment signal is not an independent discovery.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SME loan-audit text dataset and the entity/lexicon extraction that feeds the textual similarity."}],"review_version":1}