Pith. sign in

REVIEW 4 major objections 4 minor 67 references

CORE claims unified tabular anomaly detection is best formulated as in-context reconstruction: a single model, trained once on normal rows, detects anomalies in any unseen table by reconstructing each row from its most similar normal rows a

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 04:33 UTC pith:AHV3OGJT

load-bearing objection CORE's reconstruction-based unified TAD is a genuinely new formulation with consistent empirical gains, but the headline comparison borrows its OFA-TAD numbers instead of re-running them, and the paper needs code, error bars, and a trivial kNN-reconstruction baseline before the central claim is fully established. the 4 major comments →

arxiv 2607.27615 v1 pith:AHV3OGJT submitted 2026-07-30 cs.AI

CORE: In-Context Reconstruction for Unified Tabular Anomaly Detection

Yunfeng Zhao , Qingfeng Chen , Yue Tan , Shiyuan Li , Yili Wang , Yixin Liu , Shirui Pan This is my paper
classification cs.AI
keywords tabular anomaly detectionunified anomaly detectionin-context learningreconstruction-based anomaly detectionfeature alignmentdecorrelated feature selectionzero-shot generalizationoutlier detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that a single model can detect anomalies in any unseen tabular dataset by treating detection as a reconstruction problem rather than a classification problem. The proposed method, CORE, first aligns heterogeneous tables into a common feature space by greedily selecting high-variance, low-correlation original features, then scores each query row by reconstructing it from its most similar normal rows in the target dataset and measuring the reconstruction error. Because the reconstruction reference set is drawn from the target dataset itself, the model adapts to dataset-specific notions of normality without retraining and without needing labeled or synthetic anomalies. On 34 target datasets it reports the best average AUROC of 0.8488, outperforming the second-best unified method by 1.43 percentage points and winning on 26 of the 34 datasets. The result matters because it suggests that unified anomaly detection can escape the two known failure modes of classification-based approaches: loss of original feature semantics and reliance on unrepresentative synthetic outliers.

Core claim

CORE is the first reconstruction-based unified tabular anomaly detector. Its central claim is that unification is best achieved by preserving original feature semantics through decorrelated feature alignment, and that anomaly scoring should be done by in-context reconstruction: each query sample is reconstructed as a gated, weighted combination of its K nearest normal context samples, with the mean squared error as the anomaly score. The model is trained once on source datasets by minimizing reconstruction error on normal samples, then applied to arbitrary unseen target datasets using the target's own training split as context. The paper demonstrates this design on 34 datasets across diverse

What carries the argument

The key machinery has two parts. (1) Decorrelated feature alignment: for each dataset, features are scored by Var(X_j)·(1 − max_{k≠j}|Corr(X_j,X_k)|) and greedily selected to reach a unified dimension d_u, so that each selected feature is informative (high variance) yet not redundant with already-chosen features; this directly maps raw features into a shared space while keeping their semantic meaning. (2) In-context reconstruction: a shared two-layer MLP encodes context and query samples into metric, value, and gate representations; for each query, the K nearest context samples by Euclidean distance in metric space are weighted by a softmax of negative distance, fused dimension-wise with the

Load-bearing premise

The load-bearing premise is that the decorrelated feature alignment—selecting high-variance, low-correlation features (after random projection when the original dimension is smaller than d_u)—keeps the information needed to separate anomalies from normal rows; if that heuristic discards the very features that define anomalies, detection fails regardless of the reconstruction step.

What would settle it

Train or evaluate CORE on a dataset whose anomalies are defined by rare combinations of low-variance or highly correlated features (for example, a synthetic table where the anomaly is a specific conjunction of two otherwise-common binary attributes). If CORE's greedy selection drops those attributes and its AUROC falls to chance while a random-projection baseline or a full-dimension autoencoder does not, the alignment premise is falsified. A simpler check: on low-dimensional datasets (d < d_u), compare CORE with and without the random projection pre-step to see whether the projection is destro

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Unified tabular anomaly detection can be solved without binary classification, removing the need for labeled anomalies or synthetic pseudo-anomalies and avoiding the boundary-misleading artifacts they cause.
  • Because inference needs only the target dataset's training rows as context and no target-specific training, deployment cost on a new dataset is essentially zero beyond feature alignment and a forward pass.
  • Reconstruction error defined against the most similar normal rows yields dataset-specific normality, so the same frozen model can serve domains as different as healthcare, finance, and network security.
  • The scaling experiments suggest that performance continues to improve with more source data and larger models, so the approach should benefit from larger pre-training corpora.
  • The ablations show that context reconstruction is the largest contributor, implying that the framework's power lies in retrieving and aggregating relevant normal references rather than in the alignment or the gating alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An implication the authors leave implicit is that the same reconstruction-from-context machinery could be applied to graph or relational anomaly detection by swapping the feature alignment for a suitable node or edge alignment.
  • The greedy variance-correlation selection is a heuristic; a testable extension is to replace it with a learned or differentiable feature selection that optimizes end-to-end reconstruction error, which could improve low-dimensional datasets where random projection currently precedes selection.
  • The method's strong out-of-domain results hint that in-context reconstruction is learning a transferable notion of normal structure rather than dataset-specific thresholds; if true, it should also work for cross-domain drift detection, where the context set changes over time.
  • Because the anomaly score is a pure MSE in aligned space, the model could be extended to produce explanations by reporting which aligned dimensions contribute most to each query's error.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes CORE, a unified tabular anomaly detection method. A decorrelated feature alignment module greedily selects original features by variance and low inter-feature correlation to map heterogeneous datasets to a common dimension (using random projection when d < d_u). An in-context reconstruction module encodes query and context samples with a shared MLP, retrieves the K nearest context samples per query in a metric space, and reconstructs the query by gated weighted aggregation; the mean squared error between the aligned query and its reconstruction is the anomaly score. Training is performed on normal samples from 7 source datasets and evaluation is zero-shot on 34 target datasets. The paper reports an average AUROC of 0.8488, 1.43 percentage points above OFA-TAD, with wins on 26/34 datasets, together with ablations, context-ratio sensitivity, runtime, and scaling analyses. Baselines are stated to follow the results and evaluation protocol reported in OFA-TAD rather than being re-run, and no standard deviations or significance tests are reported.

Significance. If the empirical claims survive an independent re-run of the baselines, the contribution is noteworthy: CORE offers a simple reconstruction-based alternative to classification-based unified TAD, avoids synthetic anomalies, and the in-context reconstruction mechanism is conceptually clean. The 34-dataset benchmark, the ablations isolating each component, and the scaling analysis are useful and go beyond a single benchmark. However, the central quantitative claim is comparative and currently rests on external baseline numbers; with a 1.43 pp average AUROC difference and many per-dataset margins below 1 pp, the comparison must be verified in the same evaluation harness. The feature-alignment design is attractive for preserving original-feature semantics, but the random-projection path for low-dimensional datasets requires scrutiny.

major comments (4)
  1. [Experiments, Experimental Setup (Baselines)] The statement 'For baselines, we follow the results and evaluation protocol reported in OFA-TAD' indicates that the 10 baselines, including OFA-TAD, were not re-run in this paper's harness. The headline claim (Table 1: CORE 0.8488 vs OFA-TAD 0.8345; 26/34 wins) depends on per-dataset OFA-TAD scores, many of which differ by under 1 pp. Differences in splits, normalization, benchmark versions, or seeds could flip the comparison. Please re-run all baselines, especially OFA-TAD, under identical data splits and preprocessing, and report per-dataset results. This is the load-bearing condition for the central comparative claim.
  2. [Experiments, Evaluation Metrics and Implementation] The paper states 'repeat each experiment over 5 trials' but reports only point estimates in Tables 1, 4, and 5, with no standard deviations and no significance tests. For a 1.43 pp average AUROC advantage, the 26/34 per-dataset wins could be within noise. Please report mean ± std per dataset and a paired test across the 34 datasets (e.g., Wilcoxon signed-rank on per-dataset AUROC) or per-dataset bootstrap confidence intervals. Without this, the robustness of the claimed improvement is not established.
  3. [Methodology, Decorrelated Feature Alignment, last paragraph] When d < d_u, the module applies a random projection to a higher-dimensional space before performing feature selection. This is not a 'direct' alignment that preserves original-feature semantics: projected features are linear combinations of original features, and for low-dimensional datasets the projection can mix discriminative and redundant signal before selection. The w/o Align ablation replaces the entire alignment module with random projection and does not isolate the effect of the project-then-select path. Please report d_u, how many datasets satisfy d < d_u, and an ablation comparing projection-only, select-then-project, and project-then-select, or justify the current design.
  4. [Efficiency Analysis (Fig. 4 and Appendix D.3)] The runtime comparisons against OFA-TAD are meaningful only if OFA-TAD was actually executed in the same environment. Runtime cannot be copied from a prior paper in the way that AUROC numbers might be. If OFA-TAD was not re-run for Figs. 3, 4, and 8, the efficiency conclusions and the context-ratio curves for OFA-TAD are not comparable. Please clarify which OFA-TAD results are newly produced and which are taken from the original paper, and re-run the runtime and context-ratio experiments for OFA-TAD under identical conditions.
minor comments (4)
  1. [Algorithm 2 vs. equations] Algorithm 2's cross-references are off by two: 'Eq. (7)' should be Eq. (5), 'Eq. (8)' should be Eq. (6), 'Eq. (9)' should be Eq. (7), 'Eq. (10)' should be Eq. (8), and 'Eq. (11)' should be Eq. (9).
  2. [Appendix C.3, Implementation Details] The unified dimension d_u is central to the method and to the low-dimensional projection issue, but it is not listed in the hyperparameter search space or reported. Please report d_u, the temperature tau, and the exact context ratio used in training.
  3. [Fig. 3 and text] Figure 3 shows Context Nodes Ratio from 0.25 to 1.00, while the text says 'context ratio ranges from 0.1 to 1.0'. Please reconcile.
  4. [Fig. 1] Figure 1(a) and 1(b) are used to motivate the method, but no experimental details are given for the reconstruction detector, datasets, or hyperparameters used in the comparison. Please add a caption or appendix note describing how these figures were produced.

Circularity Check

0 steps flagged

No circular derivation found; the method scores anomalies by reconstruction error and is trained/evaluated without fitting to target labels.

full rationale

CORE's derivation chain contains no step in which a prediction reduces to an input by construction. The anomaly score is the MSE between an aligned query and a weighted combination of retrieved context samples (Eqs. 8-10); the model is trained only on normal source splits and applied to unseen target datasets, so no target label or anomaly score is used as a fitted input. The decorrelated feature alignment is a greedy variance/correlation feature-selection procedure (Eqs. 1-4); it is a heuristic input representation, not a parameter fitted to the predicted anomaly labels. The ablation study (removing alignment, gating, or context reconstruction) shows degradation, consistent with each component contributing independently. The only self-citation in the evaluation is the statement 'For baselines, we follow the results and evaluation protocol reported in OFA-TAD (Li et al. 2026a)'; this is a reproducibility concern about the comparative claim rather than a circular derivation, because the OFA-TAD numbers are external benchmark results from public datasets and are not constructed from CORE's own outputs or fitted values. I therefore find no significant circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

CORE does not postulate new physical or mathematical entities. Its assumptions are standard for reconstruction-based AD plus a feature-selection heuristic. The main free parameters are integer/hyperparameter choices; the paper does not report their selected values. The random projection for low-dimensional data is a modeling choice that may affect the semantic-preservation claim.

free parameters (8)
  • unified dimension d_u
    Target dimension of aligned representation; chosen by random search (Appendix C.3). When original d < d_u, a random projection is applied before feature selection.
  • number of selected context samples K
    Number of nearest context samples used for reconstruction; chosen by random search. Affects retrieval stability and model complexity.
  • temperature tau
    Softmax temperature in context weighting (Eq. 7); chosen by random search.
  • hidden layer dimension
    MLP hidden size for the shared backbone; chosen by random search.
  • metric embedding dimension p
    Output dimension of the metric head; chosen by random search.
  • number of MLP layers
    Depth of the shared backbone; chosen by random search.
  • learning rate
    Chosen by random search; affects training convergence.
  • weight decay
    Regularization coefficient; chosen by random search.
axioms (4)
  • domain assumption Context samples are normal.
    The method assumes D_k,train contains only normal samples. The paper follows the one-class setting, but many datasets listed in Table 3 have anomalies in the full data; the paper does not specify how the normal-only training subset is extracted.
  • domain assumption Reconstruction error is a valid anomaly score.
    The central scoring mechanism assumes normal samples are reconstructible from context and anomalies are not. This is the defining premise of reconstruction-based AD, not proven for tabular data.
  • domain assumption Selected features preserve anomaly-relevant semantics.
    Greedy variance/correlation selection assumes informative features are high-variance and low-redundancy. There is no guarantee that anomaly-discriminative features are retained, especially after random projection for d < d_u.
  • standard math Pearson correlation captures feature redundancy.
    The method uses linear correlation to estimate redundancy; nonlinear dependencies are ignored, which may mis-rank features.

pith-pipeline@v1.3.0-daily-deepseek · 24148 in / 11485 out tokens · 117325 ms · 2026-08-01T04:33:34.426036+00:00 · methodology

0 comments
read the original abstract

Tabular anomaly detection (TAD), which focuses on identifying abnormal samples that deviate from the majority in tabular data, has received growing attention. Recently, there has been an emerging trend towards unified TAD, which seeks to detect anomalies across different datasets using a single generalizable model. In unified TAD, aligning heterogeneous data remains challenging. While existing methods often rely on distance-based unified feature construction, they may obscure the semantics of the original features. Moreover, existing approaches typically formulate anomaly detection as a binary classification task, which may overlook diverse anomaly patterns from various datasets and be misled by unrepresentative synthetic anomalies. To address these challenges, we propose an in-COntext REconstruction approach for unified TAD (CORE for short). It introduces a decorrelated feature alignment module to directly align heterogeneous features into a unified representation space, which retains their semantic information. Meanwhile, CORE formulates unified TAD as an in-context reconstruction problem, eliminating the need for labeled or synthesized anomalies. Specifically, the in-context reconstruction module reconstructs each sample by leveraging contextual normal samples to capture dataset-specific distributions, such that reconstruction errors reflect its deviation from normality, facilitating unified TAD on arbitrary unseen datasets.

Figures

Figures reproduced from arXiv: 2607.27615 by Qingfeng Chen, Shirui Pan, Shiyuan Li, Yili Wang, Yixin Liu, Yue Tan, Yunfeng Zhao.

Figure 1
Figure 1. Figure 1: (a): Performance comparison of different feature [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall pipeline of CORE. The decorrelated feature alignment module greedily selects informative and less redundant original features and arranges them into a unified representation. Then, the in-context reconstruction module reconstructs each query according to its relevant normal context samples. The reconstruction errors serves as anomaly scores. where S denotes the set of selected features. Then, t… view at source ↗
Figure 4
Figure 4. Figure 4: Overall runtime comparison in seconds. 0.02M 0.08M 0.20M 0.60M 3.00M x:Params 0.13k 0.38k 1.2k 3.8k 7.7k y:Data 0.830 0.835 0.840 0.845 0.850 z: AUROC [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Neural scaling law analysis of CORE. tigate the scaling behavior of CORE by varying the amount of source training data and the model size. Specifically, we scale the number of source training samples from 0.13K to the full training set, and increase the number of model parameters from 0.02M to 3.00M. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: AUROC performance with varying context nodes on twelve datasets. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: AUPRC performance with varying context nodes on twelve datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Runtime comparison on twelve selected datasets. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 11 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2603.16459 , year=

    DynHD: Hallucination Detection for Diffusion Large Language Models via Denoising Dynamics Deviation Learning , author=. arXiv preprint arXiv:2603.16459 , year=

    arXiv

  2. [2]

    Multi-Stage Verification-Centric Framework for Mitigating Hallucination in Multi-Modal RAG , author=

  3. [3]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Raising the bar in graph ood generalization: Invariant learning beyond explicit environment modeling , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  4. [4]

    IEEE Transactions on Knowledge and Data Engineering , year=

    From Few-Shot to Zero-Shot: Towards Generalist Graph Anomaly Detection , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  5. [5]

    International Conference on Machine Learning , year=

    Rethinking Feature Alignment in Generalist Graph Anomaly Detection: A Relational Fingerprint-based Approach , author=. International Conference on Machine Learning , year=

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Correcting False Alarms from Unseen: Adapting Graph Anomaly Detectors at Test Time , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  7. [7]

    arXiv preprint arXiv:2606.18621 , year=

    Towards Anomaly Detection on Relational Data , author=. arXiv preprint arXiv:2606.18621 , year=

    Pith/arXiv arXiv

  8. [8]

    International Joint Conference on Artificial Intelligence , year=

    CAMERA: Adapting to Semantic Camouflage in Unsupervised Text-Attributed Graph Fraud Detection , author=. International Joint Conference on Artificial Intelligence , year=

  9. [9]

    arXiv preprint arXiv:2410.03315 , year=

    Influence-oriented personalized federated learning , author=. arXiv preprint arXiv:2410.03315 , year=

    Pith/arXiv arXiv

  10. [10]

    International Joint Conference on Artificial Intelligence , year=

    FedCIGAR: A Personalized Reconstruction Approach for Federated Graph-level Anomaly Detection , author=. International Joint Conference on Artificial Intelligence , year=

  11. [11]

    International Conference on Machine Learning , year=

    Towards One-for-All Anomaly Detection for Tabular Data , author=. International Conference on Machine Learning , year=

  12. [12]

    ACM computing surveys (CSUR) , volume=

    Deep learning for medical anomaly detection--a survey , author=. ACM computing surveys (CSUR) , volume=. 2021 , publisher=

    2021

  13. [13]

    Transactions on Emerging Telecommunications Technologies , volume=

    Network intrusion detection system: A systematic study of machine learning and deep learning approaches , author=. Transactions on Emerging Telecommunications Technologies , volume=. 2021 , publisher=

    2021

  14. [14]

    Computer Science Review , volume=

    Financial fraud detection applying data mining techniques: A comprehensive review from 2009 to 2019 , author=. Computer Science Review , volume=. 2021 , publisher=

    2009

  15. [15]

    The Twelfth International Conference on Learning Representations , year=

    Mcm: Masked cell modeling for anomaly detection in tabular data , author=. The Twelfth International Conference on Learning Representations , year=

  16. [16]

    International Conference on Learning Representations , volume=

    DRL: Decomposed representation learning for tabular anomaly detection , author=. International Conference on Learning Representations , volume=

  17. [17]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Lunar: Unifying local outlier detection methods via graph neural networks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  18. [18]

    Nature , volume=

    Accurate predictions on small data with a tabular foundation model , author=. Nature , volume=. 2025 , publisher=

    2025

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    TabDPT: Scaling Tabular Foundation Models on Real Data , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    arXiv preprint arXiv:2603.14407 , year=

    Towards One-for-All Anomaly Detection for Tabular Data , author=. arXiv preprint arXiv:2603.14407 , year=

    Pith/arXiv arXiv

  21. [21]

    International Conference on Learning Representations , year=

    UniOD: A Universal Model for Outlier Detection across Diverse Domains , author=. International Conference on Learning Representations , year=

  22. [22]

    Proceedings of the 2019 SIAM international conference on data mining , pages=

    Deep anomaly detection on attributed networks , author=. Proceedings of the 2019 SIAM international conference on data mining , pages=. 2019 , organization=

    2019

  23. [23]

    Pattern Recognition , volume=

    Reconstruction by inpainting for visual anomaly detection , author=. Pattern Recognition , volume=. 2021 , publisher=

    2021

  24. [24]

    2008 eighth ieee international conference on data mining , pages=

    Isolation forest , author=. 2008 eighth ieee international conference on data mining , pages=. 2008 , organization=

    2008

  25. [25]

    Proceedings of the 2000 ACM SIGMOD international conference on Management of data , pages=

    LOF: identifying density-based local outliers , author=. Proceedings of the 2000 ACM SIGMOD international conference on Management of data , pages=

    2000

  26. [26]

    European conference on principles of data mining and knowledge discovery , pages=

    Fast outlier detection in high dimensional spaces , author=. European conference on principles of data mining and knowledge discovery , pages=. 2002 , organization=

    2002

  27. [27]

    arXiv preprint arXiv:2007.01760 , year=

    Explainable deep one-class classification , author=. arXiv preprint arXiv:2007.01760 , year=

    Pith/arXiv arXiv 2007

  28. [28]

    2018 Wireless telecommunications symposium (WTS) , pages=

    Autoencoder-based network anomaly detection , author=. 2018 Wireless telecommunications symposium (WTS) , pages=. 2018 , organization=

    2018

  29. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Disentangling tabular data towards better one-class anomaly detection , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  30. [30]

    International conference on machine learning , pages=

    Neural transformation learning for deep anomaly detection beyond images , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  31. [31]

    1st ICML Workshop on Foundation Models for Structured Data , year=

    FoMo-0D: A Foundation Model for Zero-shot Outlier Detection , author=. 1st ICML Workshop on Foundation Models for Structured Data , year=

  32. [32]

    arXiv preprint arXiv:2602.03018 , year=

    From Zero to Hero: Advancing Zero-Shot Foundation Models for Tabular Outlier Detection , author=. arXiv preprint arXiv:2602.03018 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2603.17570 , year=

    FoMo X: Modular Explainability Signals for Outlier Detection Foundation Models , author=. arXiv preprint arXiv:2603.17570 , year=

    arXiv

  34. [34]

    arXiv preprint arXiv:2603.19497 , year=

    ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes , author=. arXiv preprint arXiv:2603.19497 , year=

    arXiv

  35. [35]

    KDD Cup Workshop for Multimodal Retrieval Augmented Generation , year=

    Multi-Stage Verification-Centric Framework for Mitigating Hallucination in Multi-Modal RAG , author=. KDD Cup Workshop for Multimodal Retrieval Augmented Generation , year=

  36. [36]

    Transactions on Graph Intelligence and Network Applications (TGINA) , year=

    From unsupervised to few-shot graph anomaly detection: A multi-scale contrastive learning approach , author=. Transactions on Graph Intelligence and Network Applications (TGINA) , year=

  37. [37]

    Li, Shiyuan and Liu, Yixin and Zheng, Yu and Li, Mei and Nguyen, Quoc Viet Hung and Pan, Shirui , booktitle=

  38. [38]

    Advances in neural information processing systems , volume=

    Adbench: Anomaly detection benchmark , author=. Advances in neural information processing systems , volume=

  39. [39]

    2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=

    Uadb: Unsupervised anomaly detection booster , author=. 2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=. 2023 , organization=

    2023

  40. [40]

    IEEE Transactions on Image Processing , volume=

    Suspicious object detection for millimeter-wave images with multi-view fusion siamese network , author=. IEEE Transactions on Image Processing , volume=. 2023 , publisher=

    2023

  41. [41]

    IEEE transactions on neural networks and learning systems , volume=

    Deep neural networks and tabular data: A survey , author=. IEEE transactions on neural networks and learning systems , volume=. 2022 , publisher=

    2022

  42. [42]

    International conference on learning representations , year=

    Anomaly detection for tabular data with internal contrastive learning , author=. International conference on learning representations , year=

  43. [43]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Data-efficient and interpretable tabular anomaly detection , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  44. [44]

    arXiv preprint arXiv:2305.15121 , year=

    Beyond individual input for deep anomaly detection on tabular data , author=. arXiv preprint arXiv:2305.15121 , year=

    Pith/arXiv arXiv

  45. [45]

    ACM computing surveys (CSUR) , volume=

    Deep learning for anomaly detection: A review , author=. ACM computing surveys (CSUR) , volume=. 2021 , publisher=

    2021

  46. [46]

    International conference on learning representations , year=

    Rapp: Novelty detection with reconstruction along projection pathway , author=. International conference on learning representations , year=

  47. [47]

    International Conference on Learning Representations , volume=

    On diffusion modeling for anomaly detection , author=. International Conference on Learning Representations , volume=

  48. [48]

    The annals of mathematical statistics , volume=

    On estimation of a probability density function and mode , author=. The annals of mathematical statistics , volume=. 1962 , publisher=

    1962

  49. [49]

    arXiv preprint arXiv:2301.04257 , year=

    ODIM: Outlier detection via likelihood of under-fitted generative models , author=. arXiv preprint arXiv:2301.04257 , year=

    Pith/arXiv arXiv

  50. [50]

    arXiv preprint arXiv:2502.05564 , year=

    Tabicl: A tabular foundation model for in-context learning on large data , author=. arXiv preprint arXiv:2502.05564 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2406.05216 , year=

    TabPFGen--Tabular Data Generation with TabPFN , author=. arXiv preprint arXiv:2406.05216 , year=

    Pith/arXiv arXiv

  52. [52]

    Proceedings of the IEEE , volume=

    A unifying review of deep and shallow anomaly detection , author=. Proceedings of the IEEE , volume=. 2021 , publisher=

    2021

  53. [53]

    Advances in neural information processing systems , volume=

    Deep anomaly detection using geometric transformations , author=. Advances in neural information processing systems , volume=

  54. [54]

    Advances in Neural Information Processing Systems , volume=

    Automatic unsupervised outlier model selection , author=. Advances in Neural Information Processing Systems , volume=

  55. [55]

    ACM computing surveys (CSUR) , volume=

    Anomaly detection: A survey , author=. ACM computing surveys (CSUR) , volume=. 2009 , publisher=

    2009

  56. [56]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  57. [57]

    Neural Computation , volume=

    A probabilistic resource allocating network for novelty detection , author=. Neural Computation , volume=. 1994 , publisher=

    1994

  58. [58]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  59. [59]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  60. [60]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    Pith/arXiv arXiv 2010

  61. [61]

    Artificial intelligence review , volume=

    A survey of outlier detection methodologies , author=. Artificial intelligence review , volume=. 2004 , publisher=

    2004

  62. [62]

    Signal processing , volume=

    A review of novelty detection , author=. Signal processing , volume=. 2014 , publisher=

    2014

  63. [63]

    Advances in neural information processing systems , volume=

    Sanflow: Semantic-aware normalizing flow for anomaly detection , author=. Advances in neural information processing systems , volume=

  64. [64]

    Neural computation , volume=

    Estimating the support of a high-dimensional distribution , author=. Neural computation , volume=. 2001 , publisher=

    2001

  65. [65]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Ecod: Unsupervised outlier detection using empirical cumulative distribution functions , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2022 , publisher=

    2022

  66. [66]

    International conference on learning representations , year=

    Deep autoencoding gaussian mixture model for unsupervised anomaly detection , author=. International conference on learning representations , year=

  67. [67]

    arXiv preprint arXiv:2005.02359 , year=

    Classification-based anomaly detection for general data , author=. arXiv preprint arXiv:2005.02359 , year=

    Pith/arXiv arXiv 2005