{"id":"c0c968b0-da4f-42ac-a70b-3ea7626fc250","arxiv_id":"2504.20862","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Tabular Data Adapters select similar public datasets, transform unlabeled private data into their format, and fuse public outlier model outputs into soft labels.","lead":"This paper introduces Tabular Data Adapters, a pipeline that finds public tabular datasets similar to an unlabeled private dataset, transforms the private data into the public format with an autoencoder, and uses pretrained outlier detectors to generate weak labels. It matters for practitioners because it attacks the cold-start labeling problem in tabular anomaly detection, where labels are scarce and public models often do not transfer directly.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dataset Transformation's preservation of per-sample outlier structure is the load-bearing assumption; Eq. 5's aggregate DS_Diff does not test it, and the paper's own ROC-AUC drop admits transformation noise.","rationale":"The reader's weakest_assumption correctly identifies the unvalidated link between Dataset Transformation and label quality. I agree with that identification. I considered alternative concerns—missing error bars, unspecified m/n, no released code, and the textual contradiction that Method2's PR-AUC exceeds the supervised Best OD despite §5.1 saying otherwise—but these are secondary or resolvable. The transformation-preservation issue is the one that, if wrong, makes the central claim unsupported. The empirical Table 1 provides partial reassurance (ROC-AUC 0.689/0.683 is above chance), so this is not a proven fatal flaw; it is an unverified load-bearing premise. The proposed test uses the already-available evaluation labels to measure per-sample separation on crossover data and compares against a nontrivial control. If the test passes, the concern is resolved and the conditional accept stands; if it fails, the soft-label gains should be attributed to the voting/fusion scheme rather than to structure-preserving transfer and the paper would need major revision. Since the concern is addressable, I keep the reader's CONDITIONAL verdict (UNCHANGED).","tokens_in":14658,"tokens_out":10459,"duration_ms":125207,"concrete_test":"Use the known private labels that are available in the leave-one-out evaluation (but not during soft-label generation) to compute, for every held-out dataset, the ROC-AUC of the best public model's scores on crossover samples restricted to true outliers versus inliers. Compare this with a control that replaces the learned transformation (Encoderprv/mtheta/Decoderpub) with a fixed per-feature quantile transform plus zero-padding to the public feature dimension, keeping the same public models and voting. If the learned transformation does not substantially improve per-sample outlier/inlier separation over the control, then Eq. 5's aggregate DS_Diff cannot certify that outlier structure transfers, and the reported soft-label gains should not be attributed to the transformation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that soft labels are more accurate than unsupervised baselines; the only route from private features to public-model scores is the Dataset Transformation (Eq. 4). The load-bearing assumption is that this learned mapping preserves the per-sample outlier structure of the private data. The paper's validation, Eq. 5, compares only aggregate DS_Diff values between datasets; DS_Diff is not defined in this paper, no DS_Diff numbers are reported, and the comparison says nothing about whether the outliers in Xprv remain outliers after crossover. The training loss (Eq. 3) is plain MSE reconstruction on mixed batches, so rare outliers contribute negligibly; there is no term that encourages preserving outlier-specific directions. The paper itself (§5.1) attributes the lower ROC-AUC to 'inherent noise introduced during dataset transformation,' which is exactly the failure mode. If the transformation smooths or repositions anomalous private samples into the public inlier manifold, then even a perfectly similar public dataset yields near-random labels, and the Balanced Accuracy/F1 gains in Table 1 would have to be explained by fusion or threshold effects rather than by faithful label transfer. This is not contradicted by the aggregate results because Table 1 reports no per-sample separation on crossover data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Tabular Data Adapters (TDA), a pipeline for generating soft labels for unlabeled private tabular data in outlier detection. A PCA-based reconstruction-error curve is used to rank public datasets by statistical similarity to the private dataset; a shared-latent autoencoder pair (the Dataset Transformation module) then maps private samples into the selected public dataset's feature space; publicly pre-trained outlier-detection models are applied to these 'crossover' samples, and their predictions are fused by majority voting. Two variants are introduced: Top1-DS, which uses the single most similar public dataset with its m best models, and TopN-DS, which uses the top n public datasets with one best model each. The method is evaluated in a leave-one-out fashion over 50 tabular datasets from ADBench and three financial datasets, with Table 1 reporting balanced accuracy, F1, PR-AUC, ROC-AUC, and runtime against three baselines: Best OD (supervised upper bound), Avg OD, and Default OD.","tokens_in":14890,"tokens_out":3772,"duration_ms":43075,"significance":"If the central claim holds, the contribution is practically useful: it offers a cold-start path to weak labels for tabular outlier detection without manual labeling, and the reported compute savings (9.94 and 27.28 minutes versus 282.64 for Avg OD) are substantial. The paper deserves credit for a broad 50-dataset leave-one-out evaluation, two clearly described method variants, a domain-level ablation, and an explicit analysis of where computation time is spent. However, the current evidence does not yet establish the load-bearing claim that the Dataset Transformation preserves the per-sample outlier structure needed for meaningful label transfer. The missing validation of that step, combined with the unresolved ROC-AUC shortfall and under-specified hyperparameters, means the headline 'more accurate' claim is not yet substantiated as written.","major_comments":[{"comment":"The success criterion for Dataset Transformation is the inequality DS_Diff(Xprv, Xpub) > DS_Diff(Xprv, Xhat_co), but DS_Diff is never defined in this manuscript, no numerical DS_Diff values are reported, and the criterion is aggregate: it compares reconstruction-error similarity between whole datasets, not whether individual private outliers remain outliers after crossover. Because Eq. (3) is plain MSE reconstruction on mixed batches, rare outliers contribute negligibly to the training loss, so there is no mechanism that explicitly preserves outlier-specific directions. This is load-bearing: without evidence that outlier/inlier separability survives the crossover mapping, the public models' scores on crossover data may be near-random even when the selected public dataset is highly similar. I ask the authors to provide a direct per-sample validation, e.g., reporting, for each held-out private dataset, the ROC-AUC or rank correlation between the public model's crossover scores and the private ground-truth labels, or an outlier-score distribution comparison before and after transformation. The paper's own §5.1 remark that the ROC-AUC drop is due to 'inherent noise introduced during dataset transformation' makes this test necessary rather than optional.","section":"Section 3.1, Eq. (5)"},{"comment":"The abstract and Section 5.1 state that the method provides 'more accurate annotations' than baseline approaches, but Table 1 shows both methods have lower ROC-AUC (0.689 and 0.683) than Avg OD (0.691) and Default OD (0.690). Since ROC-AUC is a standard threshold-independent metric for outlier-detection quality, the unqualified 'more accurate' claim is contradicted on one of the four reported metrics. The authors should either qualify the claim to the metrics where the method wins, or provide an explanation and supporting analysis for the ROC-AUC shortfall. Additionally, no standard deviations, confidence intervals, significance tests, or number of random seeds are reported, so it is unclear whether the differences in balanced accuracy, F1, and PR-AUC are robust across runs and dataset subgroups.","section":"Table 1 and Abstract"},{"comment":"The parameters m (number of best public models in Method1) and n (number of top public datasets in Method2) are never given concrete values for the experiments in Table 1, even though they are the key free parameters of the method. Figure 5 plots m and n up to 9 or 10, but the text does not state which values were used for the headline results. Similarly, the Dataset Similarity Measure is described as varying the number of principal components 'from 1 to 100,' but several datasets in Table 2 have fewer than 100 features; the handling of this boundary case is not specified. Without these details the experiments are not reproducible and the sensitivity of the reported gains to m and n cannot be assessed.","section":"Sections 3.2 and 4.3"},{"comment":"The text says 'To find the most similar public datasets, we rank them descending by their SAD to the private dataset,' which is inconsistent with the definition of SAD as a sum of absolute differences where lower values indicate higher similarity. Figure 8 plots a 'similarity score' between 0.2 and 1.0 but does not define how that score is derived from SAD. Because the selection of the public dataset is the first step of the pipeline, this ambiguity should be resolved by defining the similarity score precisely and correcting the ranking direction.","section":"Section 3.1, Dataset Similarity Measure"},{"comment":"The evaluation metrics include PR-AUC and ROC-AUC, but the description of the method says soft labels are produced by majority voting, which yields binary predictions rather than a continuous score. It is not specified whether PR-AUC and ROC-AUC are computed from vote proportions, from averaged anomaly scores of the individual public models, or from some other ranking. This matters because the method's PR-AUC gain over the baselines could in part reflect how the score is constructed. The authors should state exactly what quantity is used to rank samples for these two metrics.","section":"Section 4.3 and Table 1"},{"comment":"The compute-time comparison, while informative, does not account for the cost of training the Dataset Transformation model for each new private dataset, nor for the presumed pre-computation of the public model repository. The claim of 'reducing computational time' is relative to baselines that run all 11 OD algorithms with multiple hyperparameter configurations, which is a fair baseline for a cold-start scenario, but the paper should state whether the reported Method1/Method2 times include transformation training and model inference, and should clarify what is assumed to be available as pre-existing assets.","section":"Section 5.2 and Figure 5"},{"comment":"The box plots in Figure 4 are described as showing 'consistent superiority' and 'comparable performance,' but the ROC-AUC panel shows the methods slightly below the baselines on average, and no statistical testing is reported. The visual claim of consistency should be backed by per-dataset counts of wins/losses against each baseline for each metric, or by paired tests; otherwise the figure overstates the strength of the evidence.","section":"Section 5.1, Figure 4"}],"minor_comments":[{"comment":"There are several typographical issues, including 'V esselin' in the author list, 'Leave one out' used as a compound adjective, and inconsistent hyphenation of 'leave-one-out'; these should be cleaned up in revision.","section":"Throughout"},{"comment":"The loss notation uses nprv and npub before defining them; please state explicitly that these are the batch sizes for the private and public batches, respectively.","section":"Section 3.1, Eq. (3)"},{"comment":"The text says 47 datasets were selected from ADBench and 10 were excluded, but ADBench is stated to contain 57 datasets; please verify the arithmetic and clarify whether the three additional financial datasets are counted within the 50 used in Table 1.","section":"Section 4.1"},{"comment":"The caption says the y-axis shows 'Mean Reconstruction Error' but the figure is described as a similarity plot; please ensure the axes and the SAD computation are described consistently in the caption and the text.","section":"Figure 2 caption"},{"comment":"Reference [14] is the self-cited source for Dataset Transformation and is central to the method, but the paper does not summarize the DS_Diff metric from that work; a brief definition in the text or appendix would make this manuscript self-contained.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the manuscript's core idea is timely and the experimental scope is commendable, but the central validation gap concerning Dataset Transformation (Eq. 5, Section 3.1) is serious enough that I cannot recommend acceptance without the authors demonstrating per-sample outlier-structure preservation and clarifying the reporting of m, n, and the ROC-AUC discrepancy. The repeated reliance on self-cited [14] for a load-bearing component without independent validation in this paper is a reviewer concern, though not by itself a reason to reject. The paper fits the cs.LG scope and, with the requested experiments and reporting fixes, could become a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a genuine end-to-end contribution: a PCA-based dataset similarity ranker feeding a self-supervised crossover transformation that lets public OD models score private unlabeled data. Second, the headline claim of \"more accurate soft labels\" is only partially supported: Method1 and Method2 beat the unsupervised baselines on balanced accuracy, F1, and PR-AUC, but both are worse on ROC-AUC, and the paper acknowledges this.\n\nWhat's new: the combination of the similarity measure with Dataset Transformation (the authors' own prior work) is not in the cited literature, and the leave-one-out evaluation across 50 ADBench datasets is a reasonable amount of evidence. The compute savings are large — about 10 minutes vs 280 for Avg OD — and Figure 8's upward trend between similarity and balanced accuracy is a useful sanity check that the similarity measure carries signal. Credit where due: the paper is honestly written, openly cites its own prior module, and discusses the ROC-AUC weakness rather than hiding it.\n\nSoft spots, in order of severity. First, the load-bearing assumption: the transformation (Eq. 4) is trained with plain MSE reconstruction on mixed batches, and nothing in the loss preserves per-sample outlier structure. The paper validates the transformation only via the aggregate DS Diff metric (Eq. 5), which is not defined here and for which no numbers are reported. The authors attribute the ROC-AUC drop to \"inherent noise introduced during dataset transformation,\" which is exactly the failure mode the stress-test flags. This is a real gap. Second, there are no error bars, significance tests, or seeds; the gains on balanced accuracy (+0.06) and F1 (+0.05) are modest, and variance could change the story. Third, m and n are never stated for the main results. Fourth, no code or data is released yet.\n\nThe circularity concern is real but moderate. The transformation module is self-cited and the evaluation depends on it, but the prior work is peer-reviewed, so this is not disqualifying.\n\nBottom line: the central argument mostly holds for balanced accuracy, F1, and PR-AUC, but the abstract's \"more accurate\" is overstated when ROC-AUC is lower. The paper is a solid practical pipeline for cold-start labeling, and a serious referee could push it into publishable shape by requiring per-sample transfer analysis, error bars, and explicit hyperparameters. I would send it to review with major revision.","headline":"A practical soft-labeling pipeline with real gains on three of four metrics; the key transfer assumption is under-validated and the abstract overclaims, but it deserves a serious referee.","tokens_in":15445,"tokens_out":2425,"would_cite":true,"duration_ms":24400,"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 claims that unlabeled private tabular data can be soft-labeled for outlier detection by finding a statistically similar public dataset, transforming the private rows into that dataset's format with a shared-latent autoencoder…","keywords":["outlier detection","soft labels","tabular data","weak supervision","dataset transformation","autoencoder","dataset similarity","cold start"],"falsifier":"Take a private dataset whose outliers are defined by a rare combination of two features, train the Dataset Transformation autoencoder, and check whether TDA's soft labels track the true outliers; if the crossover projection compresses those two features into one, balanced accuracy should fall to chance while the DS Diff inequality still holds, and that outcome would falsify the claim that successful format alignment preserves outlier structure.","tokens_in":1673,"feed_emoji":"📊","tokens_out":2678,"duration_ms":87106,"temperature":0.7,"pith_summary":"Tabular Data Adapters is a method for labeling, without manual annotation, the outlier versus normal status of rows in a private tabular dataset. The idea is to find a public dataset that is statistically similar, train a shared autoencoder that rewrites private rows into the public dataset's format, and then run that public dataset's tuned outlier detection models on the rewritten rows. The paper argues that this generates more accurate soft labels than running standard unsupervised outlier detectors directly on the private data, while using far less computation time. If true, the method lets data owners escape the cold-start problem by bootstrapping downstream models from weak labels produced by public models instead of paying for labeling.","feed_headline":"No labels needed: outlier detection borrows similar public datasets","feed_subtitle":"A shared autoencoder maps private rows into a public format so tuned models can flag outliers in minutes.","key_machinery":"The load-bearing object is the Dataset Transformation autoencoder: two dataset-specific encoder-decoder pairs joined by shared layers $m_\\theta$. Training alternates private and public batches and minimizes a reconstruction loss (Equations 1-3); inference produces crossover samples $\\hat{x}^{co} = \\mathrm{Dec}_{pub}(m_\\theta(\\mathrm{Enc}_{prv}(x^{prv})))$ (Equation 4), which put private rows into the public format so a model trained on the public dataset can score them. Its success is checked only by the DS Diff inequality (Equation 5), which compares aggregate PCA reconstruction-error curves before and after transformation. The companion Dataset Similarity Measure ranks public datasets by the sum of absolute differences between the private and public reconstruction-error curves, and the final soft labels come from majority voting over the selected public models' predictions on the crossover samples.","core_discovery":"The paper's central claim is that you can generate useful soft labels for a private, unlabeled tabular dataset by ranking public tabular datasets with a PCA reconstruction-error similarity measure, training a shared-latent autoencoder that rewrites private samples in the chosen public dataset's format, and then applying that public dataset's already-tuned outlier detection models to the rewritten samples and fusing their votes. In a leave-one-out study over 50 tabular datasets, the authors report that both variants, Top1-DS and TopN-DS, beat the averaged and default unsupervised baselines on balanced accuracy, F1-score, and PR-AUC while using far less compute: Top1-DS averages 0.639 balanced accuracy and 9.94 minutes, versus 0.578 and 282.64 minutes for the averaged baseline. They also report that ROC-AUC remains at baseline level and that both methods stay below the supervised Best OD upper bound, which assumes labels that the method is designed to avoid needing.","pith_inferences":["The paper does not test whether the soft labels actually improve a downstream classifier trained on the private data; a natural next experiment is to train a model on TDA-soft-labeled private rows and compare it against models trained on the baseline labels or on unlabeled data with self-training.","Because the similarity measure only compares reconstruction-error curves, it could rank two datasets as similar even when their outlier-generating mechanisms differ; a sharper similarity test could compare per-sample reconstruction patterns or use a held-out probe of public outlier labels.","If the transformation is faithful enough to preserve outlier structure, the same pipeline could transfer other public-model capabilities to private tabular data, such as imputation or concept-drift detection, not just outlier flags.","The privacy angle is implicit: the private data never needs to be shared, but the trained transformation model is itself a function of private data, and whether crossover samples or the shared layers leak private information is not analyzed."],"forward_implications":["If the claim is right, an organization with an unlabeled private table can obtain outlier and normal weak labels without manual annotation, by pointing the pipeline at a repository of public datasets with tuned outlier detection models.","Method1's 9.94-minute average runtime, versus 282.64 minutes for the averaged baseline, makes per-dataset weak labeling cheap enough to rerun as private data shifts over time.","The reported upward trend between dataset similarity and balanced accuracy implies that maintaining a diverse index of public datasets directly improves label quality, so the similarity ranking is not a detail but a driver of performance.","Because ROC-AUC stays at baseline level while F1 and PR-AUC improve, the method's gains are concentrated in precision-oriented metrics, meaning it is most useful when outlier flags are used to prioritize candidates rather than to produce exact counts.","Both the single-dataset variant and the multi-dataset fusion variant beat the unsupervised baselines, so the approach does not depend on knowing in advance which public dataset is the perfect match."],"supporting_citations":[{"why":"Supplies the Dataset Transformation autoencoder with private and public encoder-decoder pairs and shared layers, plus the DS Diff similarity check that the soft-labeling pipeline reuses.","marker":"[14]"},{"why":"Supplies most of the 50 evaluation datasets and the 11 unsupervised outlier detection algorithms used both as baselines and as the source of each public dataset's tuned models.","marker":"[11]"},{"why":"Defines the comparative-evaluation protocol of averaging each algorithm over multiple hyperparameter configurations, which the Avg OD baseline follows.","marker":"[10]"},{"why":"Provides the PCA reconstruction-error basis for the Dataset Similarity Measure that ranks public datasets by similarity to the private dataset.","marker":"[1]"}],"fun_headline_variants":["Outlier detection without labels by adapting to public data","Shared autoencoder turns private rows public for outlier flags","Label-free outlier detection via public dataset adaptation","PCA picks similar public set, encoder rewrites private rows","Soft labels from similar public datasets speed outlier detection"],"cache_read_input_tokens":17536,"weakest_assumption_plain":"The one premise everything rests on is that the neural network that compresses private rows and rebuilds them in the public format keeps whatever makes private outliers anomalous; if that reshaping smooths away the anomaly, the labels are random no matter how similar the chosen public dataset looks.","fun_headline_variants_meta":{"raw":{"variants":["Outlier detection without labels by adapting to public data","Shared autoencoder turns private rows public for outlier flags","Label-free outlier detection via public dataset adaptation","PCA picks similar public set, encoder rewrites private rows","Soft labels from similar public datasets speed outlier detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1602,"prompt_tokens":921,"completion_tokens":681,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":607}},"tokens_in":537,"tokens_out":681,"duration_ms":7056,"temperature":1.0,"reasoning_tokens":607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:17:22.389212+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a private dataset whose outliers are defined by a rare combination of two features, train the Dataset Transformation autoencoder, and check whether TDA's soft labels track the true outliers; if the crossover projection compresses those two features into one, balanced accuracy should fall to chance while the DS Diff inequality still holds, and that outcome would falsify the claim that successful format alignment preserves outlier structure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies most of the 50 evaluation datasets and the 11 unsupervised outlier detection algorithms used both as baselines and as the source of each public dataset's tuned models."},{"cited_title":"Abdi and L","cited_arxiv_id":null,"evidence_quote":"Provides the PCA reconstruction-error basis for the Dataset Similarity Measure that ranks public datasets by similarity to the private dataset."}],"review_version":1}