CL-ISR: A Contrastive Learning and Implicit Stance Reasoning Framework for Misleading Text Detection on Social Media

Chia-En Chiang; Cuiqianhe Du; Tianyi Huang; Zikun Cui

REVIEW 3 major objections 5 minor 1 cited by

CL-ISR: A Contrastive Learning and Implicit Stance Reasoning Framework for Misleading Text Detection on Social Media

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that jointly training contrastive learning and implicit stance reasoning lifts misleading-text detection F1 to 91.7 on FakeNewsNet, 86.3 on PHEME, and 84.8 on Weibo-Misinfo, beating BERT, RoBERTa, and SVM.

desk verdict Reasonable combination of standard ideas, but the ISR loss is undefined on two of three datasets, so the headline results aren't attributable to the model as described. read the letter →

arxiv 2506.05107 v1 pith:SVTDKRIR submitted 2025-06-05 cs.CL

Tianyi Huang , Zikun Cui , Cuiqianhe Du , Chia-En Chiang This is my paper

classification cs.CL

keywords misleadingtextdetectioncontrastivelearningimplicitstancereasoningattention-basedfusionInfoNCElosscross-domaintransfer

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that misleading-text detection on social media improves when two learning signals are trained together: contrastive learning on augmented pairs of posts, and an auxiliary implicit-stance-reasoning task that predicts whether a post supports, opposes, or stays neutral toward a topic. The CL-ISR framework fuses the two feature representations with a learned attention gate and optimizes a weighted sum of the contrastive loss, the stance cross-entropy, and the detection cross-entropy. The authors report F1 scores of 91.7 on FakeNewsNet, 86.3 on PHEME, and 84.8 on Weibo-Misinfo, above the SVM, BERT, and RoBERTa baselines on all three datasets. Their ablation attributes 5.0 F1 points to contrastive learning, 3.8 to stance reasoning, and 2.8 to the fusion gate. If these results hold, the recipe offers a practical way to strengthen misinformation filters, particularly for posts that mislead through stance shifts or emotional wording.

What carries the argument

The load-bearing mechanism is the attention-gated fusion of two learned representations: contrastive features $h_i$ from an encoder trained with InfoNCE on augmented views, and stance features $s_i$ from a BiLSTM-plus-attention encoder trained with three-way cross-entropy (oppose, neutral, support). The gate $\beta_i = \sigma(W_h h_i + W_s s_i + b)$ in Eq. (13) forms a per-sample convex combination $r_i = \beta_i h_i + (1 - \beta_i) s_i$, letting the model lean on semantic distinctiveness or on stance information depending on the post. The total objective $\mathcal{L}_{\text{total}} = \alpha_1 \mathcal{L}_{CL} + \alpha_2 \mathcal{L}_{ISR} + \alpha_3 \mathcal{L}_{class} + \mathcal{L}_{reg}$ ties the modules together, and the paper's consistent conclusion is that the gate and both losses are needed for the reported F1 scores.

What would settle it

Inspect the FakeNewsNet and Weibo-Misinfo annotation files for stance labels in $\{-1,0,+1\}$; if neither dataset contains them, Eq. (11)'s ISR loss is undefined on those data, so the Table 1 scores could not come from the described training objective. A complementary test is to rerun the Table 2 ablation with $\alpha_2 = 0$ on those two datasets and check whether the F1 drop is the 3.8-point decline the paper attributes to stance reasoning.

Watch

Extended reading notes

Core claim

On its own terms, the central discovery is that the joint objective in Eq. (16) — an InfoNCE contrastive loss combined with an implicit-stance classification loss and a binary misleading-versus-truthful loss — produces a detector more accurate than any single component or standard fine-tuning. Table 1 reports CL-ISR at 91.7 F1 on FakeNewsNet (4.8 points above RoBERTa), 86.3 on PHEME, and 84.8 on Weibo-Misinfo (6.4 above BERT). Table 2 reports that each removal degrades accuracy and recall, with contrastive learning the largest contributor and stance reasoning the main balancer of false positives against false negatives. The paper also reports cross-domain transfer gains, for example 84.1 F1 when moving from FakeNewsNet to PHEME, again above BERT and RoBERTa.

Load-bearing premise

The framework depends on having stance labels $y_i \in \{-1,0,+1\}$ for every dataset used to train the ISR loss, yet the paper describes manual stance labels only for PHEME; if FakeNewsNet and Weibo-Misinfo lack such annotations, the objective cannot be implemented as written and the reported F1 gains cannot be attributed to the stance module.

Editorial extensions

If this is right

Misleading-text detectors can be improved by adding an auxiliary stance-prediction task even when the final output is a binary true-versus-misleading label.
Contrastive learning with hybrid augmentation (random deletion plus synonym replacement) is the largest single contributor, especially to recall, which drops by 5.8 points when the module is removed.
The three components are complementary: removing CL costs 5.0 F1 points, removing ISR costs 3.8, and removing the fusion gate costs 2.8, so future systems should keep both signals rather than choose one.
The model transfers across languages and platforms better than fine-tuned baselines, reporting 84.1 F1 on the FakeNewsNet-to-PHEME shift and 78.6 on PHEME-to-Weibo-Misinfo, suggesting less domain-bound representations.
Since the framework uses BERT-family tokenizers and reaches 84.8 on a Chinese dataset, the same architecture could be applied to other platforms or low-resource languages.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A reader could test whether gold stance labels are necessary by replacing them with auto-generated stance predictions on datasets without annotations; if the 3.8-point ISR contribution survives, the module's value is in the stance-shaped inductive bias rather than annotation quality.
The learned fusion gate $\beta_i$ could be turned into an interpretability tool: plotting its distribution across posts would reveal which texts the model decides are best read through stance rather than semantic contrast, something the paper does not itself report.
Because PHEME's stance labels were converted into the binary misleading labels by the authors, the stance supervision and the detection target share provenance there; an independent dataset with separately collected stance and veracity labels would cleanly separate the two contributions.
The same architecture could be tested on dedicated rumor or fact-check benchmarks with independent veracity labels to see whether the stance prior helps when the detection label is not derived from stance.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First thing to know: this is a fairly standard stack—InfoNCE contrastive loss, a BiLSTM stance encoder with attention, and a weighted fusion layer—applied to three misinformation datasets. The combination is not crazy, and the paper is readable. But the central claim falls apart on inspection: the implicit stance reasoning loss in Eq. (11) requires per-text stance labels y_i in {-1,0,+1}, and Section IV.A only says PHEME has stance annotations. For FakeNewsNet and Weibo-Misinfo, the paper never says where those labels come from. That means the total loss in Eq. (16) has an undefined term for two of the three datasets, so the reported F1 improvements on those datasets cannot be credited to the full CL-ISR model. This is not a missing-code issue; the method description itself is incomplete.

What's actually new: not much. Each component is standard, and similar contrastive and stance-based detectors are cited. The only new thing is the specific combination, which is a legitimate but modest contribution. The authors do run cross-domain experiments, which is a plus, and the attention-based fusion of CL and ISR features is a reasonable design choice. The ablation story, despite presentation problems, is internally consistent in direction.

Soft spots: the missing stance labels are the load-bearing one. Beyond that, the paper reports no error bars, even though it says results are averaged over three runs. The conclusion's 'only CL gives 88.1, only ISR gives 87.3' doesn't obviously match Table 2, which reports drops rather than absolute numbers. There is also a 'five datasets' typo when only three are used. None of these by themselves are fatal, but they suggest the experimental record wasn't given the same care as the model description.

Who it's for: a reader wanting a quick recipe for adding contrastive learning and a stance module to a misinformation classifier might find the architecture description useful. But anyone comparing methods needs the missing supervision details and reliable numbers first.

Recommendation: I would not send this to peer review as is. The central method is underspecified to the point of not being implementable on two of the three datasets. If the authors clarify how stance labels are obtained (or restrict the ISR loss to PHEME), add error bars, and fix the ablation inconsistencies, it could be a modest workshop-level contribution. For now, desk reject.

Referee Report

3 major / 5 minor

Summary. The paper proposes CL-ISR, a framework that combines contrastive learning with an implicit stance reasoning module for detecting misleading social media text. It describes a two-module architecture with an attention-based fusion layer and a combined loss comprising contrastive, stance-reasoning, and classification terms. The paper reports experiments on FakeNewsNet, PHEME, and Weibo-Misinfo, claiming F1 improvements over SVM, BERT, and RoBERTa baselines, along with ablation and cross-domain generalization results. The central claim, stated in Section V, is that CL-ISR 'always outperforms them on all public datasets.'

Significance. If the reported results were fully supported, the framework would be a reasonable engineering contribution: it addresses a practically important task and combines two known techniques (contrastive learning and stance modeling) in a straightforward way. The paper also reports useful experimental environment details and uses three public datasets of different languages. However, the manuscript as written does not establish the central claim. The most serious problem is that the implicit stance reasoning loss in Eq. (16) requires stance labels that are only described for PHEME, leaving the training objective undefined for FakeNewsNet and Weibo-Misinfo. In addition, the ablation and cross-domain numbers are internally inconsistent and no uncertainty estimates are provided. These issues make the reported F1 gains not defensible as stated.

major comments (3)

[Section III.B, III.C, and IV.A] The overall objective in Eq. (16) includes the term alpha_2 * L_ISR, and L_ISR in Eq. (11) requires stance labels y_i in {-1,0,+1} as defined in Section III.B. Section IV.A explicitly mentions stance annotations only for PHEME: 'All tweets in PHEME have been manually marked the stance.' For FakeNewsNet and Weibo-Misinfo, the paper describes only real/fake labels or content labels and provides no stance annotation procedure, no automatic stance inference method, and no external stance label source. Consequently, on those two datasets the term alpha_2 * L_ISR has no defined supervision, the training objective in Eq. (16) cannot be evaluated as written, and the Table 1 F1 results for FakeNewsNet and Weibo-Misinfo cannot be attributed to the full CL-ISR architecture. This directly undermines the 'always outperforms' claim in Section V.
[Section V and Table 2] The conclusion states that using only contrastive learning gives an F1 score of 88.1% and using only implicit stance reasoning gives 87.3%, but Table 2 reports only relative declines (e.g., -5.0, -3.8, -2.8) without the base F1 values; there is no row in Table 2 from which the 88.1% and 87.3% figures can be derived. Since Section IV.B says final metrics were averaged over three runs, the absence of any variance or significance measure also makes it impossible to judge whether the 4-5 point differences in Table 1 are meaningful or within run-to-run variation.
[Section IV.B, IV.C, and Table 3] The data augmentation strategy and the loss weights alpha_1, alpha_2, alpha_3 are selected using the same experiments that produce the final reported results. For example, Table 3 compares augmentation strategies and selects the 'hybrid enhancement' strategy that yields the best F1, but no nested validation or separate hold-out protocol is described for these choices. The reported gains therefore reflect, at least in part, in-sample tuning rather than an independent confirmation of the framework. A fixed configuration chosen on validation data only, or a nested evaluation, would be needed to support the claim that CL-ISR always outperforms baselines.

minor comments (5)

[Section IV.A] The text says 'The application of these five datasets promoted the cross-media evaluation,' but only three datasets are used in the paper; this should read 'three datasets.'
[Section III.B, Eq. (6)] Eq. (6) uses y_ic as if it is a one-hot indicator while the text defines y_i as a scalar in {-1,0,+1}. Please clarify the notation: either define y_ic explicitly as one-hot encoding of the stance label, or rewrite the cross-entropy formula accordingly.
[Table 1 and Section IV.C] Table 1 is titled 'Cross-domain detection performance' but reports in-domain evaluation on each dataset; the cross-domain results appear only in Figure 2. The title and the text describing Table 1 should be aligned.
[Section IV.B and Section IV.C] The paper states that final metrics are averaged over three independent runs, but no standard deviations or confidence intervals are reported anywhere. Please include variance estimates for the main results and ablation experiments.
[References] References [5] and [14] appear to describe the same paper (Deng/Sun et al., 'Implicit Stance Detection with Hashtag Semantic Enrichment') under two different reference numbers; please merge or distinguish them properly.

Circularity Check

1 steps flagged · score 6.0 of 10

Partial circularity in the PHEME evaluation: the binary misleading label is derived from the same stance labels that supervise the ISR module, so the claimed ISR benefit on that dataset is built into the label construction; the two other datasets have a non-circular but severe missing-supervision gap.

self definitional [Section IV.A (PHEME description), Section III.B Eqs. (6)/(11), Section III.C Eq. (15)]
"All tweets in PHEME have been manually marked the stance, and we further marked these stance labels into binary labels, i.e., misleading and non-misleading. ... y_i ∈ {−1,0, +1} is its stance label (-1 for opposition, 0 for neutrality, +1 for support)."

For PHEME, the classification target ('misleading vs non-misleading') is stated to be obtained by converting the stance labels into binary labels. The ISR loss L_ISR (Eqs. 6, 11) is supervised by exactly those stance labels, while the classification loss L_class (Eq. 15) is supervised by the binary labels derived from them. The two supervision terms in Eq. (16) therefore encode the same underlying annotation. Attributing the PHEME performance gain to implicit stance reasoning is thus partly guaranteed by the label-construction rule rather than by detecting an independently defined misleading property. An independent evaluation would require the misleading label to be annotated separately from the stance labels used to train the ISR module.

full rationale

Most of the paper is a standard combination of InfoNCE contrastive loss, a BiLSTM+attention stance module, and a fused classifier; there is no derivation chain in which a fitted parameter is renamed as a prediction, and the external baselines provide independent content. The one genuinely circular-by-construction element is the PHEME dataset description: because the binary misleading label is made by converting the stance labels, using those stance labels as ISR supervision and then attributing PHEME gains to ISR reduces to exploiting the label-generation rule rather than to detecting independently defined misleading content. The paper also has a severe missing-support problem for FakeNewsNet and Weibo-Misinfo: L_ISR in Eq. (11) requires stance labels, but Section IV.A only describes stance labels for PHEME; without an additional stance-labeling procedure or a transferred stance encoder, the total loss in Eq. (16) cannot be computed on the other two datasets, so those F1 gains are not attributable to the full architecture as written. That is a correctness/attribution gap, not a circular reduction, and is not counted in the circularity score. Self-citations such as references [5] and [14] are background and are not load-bearing, so they do not further raise the score.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new scientific entities or derivations; it reports an empirical model. The central claim depends on a set of hand-set hyperparameters and, critically, on unstated stance supervision for two of three datasets.

free parameters (5)

Loss weighting coefficients alpha_1, alpha_2, alpha_3 = not reported; tuned on validation set
Determined through tuning on the validation set (Section IV.B); chosen to maximize reported performance.
Temperature tau = 0.07
Set manually; standard contrastive learning temperature that controls the reported F1.
L2 regularization coefficient lambda = 1e-4
Set manually in Section IV.B; affects optimization landscape.
Data augmentation strategy = hybrid deletion + synonym replacement
Selected because Table 3 shows the highest F1 among augmentation options; chosen post hoc from the reported experiments.
FakeNewsNet subset size and selection = about 15,000 randomly selected tweets
Subset and random selection process not specified, yet all FakeNewsNet results depend on it.

assumptions (4)

ad hoc to paper PHEME stance labels can be converted into binary misleading/non-misleading labels without additional validation.
Section IV.A states 'we further marked these stance labels into binary labels', but gives no annotation protocol or quality check.
ad hoc to paper Stance supervision is available or inferable for FakeNewsNet and Weibo-Misinfo.
Required for Eq. (11) but never described; without it the ISR module cannot be trained on these datasets.
domain assumption Data augmentation preserves semantic category (truthful vs misleading).
Random deletion/insertion/replacement used to build positive pairs in Eq. (2); if augmentation flips the label, contrastive learning is mis-supervised.
domain assumption The benchmark labels in the public datasets are ground truth.
Standard assumption when using FakeNewsNet, PHEME, and Weibo-Misinfo.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CL-ISR: A Contrastive Learning and Implicit Stance Reasoning Framework for Misleading Text Detection on Social Media." pith.science (2026). https://pith.science/paper/SVTDKRIR

@misc{pith2026250605107,
  author       = {Pith},
  title        = {Pith review of: CL-ISR: A Contrastive Learning and Implicit Stance Reasoning Framework for Misleading Text Detection on Social Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SVTDKRIR}},
  note         = {Machine review of arXiv:2506.05107}
}

read the original abstract

Misleading text detection on social media platforms is a critical research area, as these texts can lead to public misunderstanding, social panic and even economic losses. This paper proposes a novel framework - CL-ISR (Contrastive Learning and Implicit Stance Reasoning), which combines contrastive learning and implicit stance reasoning, to improve the detection accuracy of misleading texts on social media. First, we use the contrastive learning algorithm to improve the model's learning ability of semantic differences between truthful and misleading texts. Contrastive learning could help the model to better capture the distinguishing features between different categories by constructing positive and negative sample pairs. This approach enables the model to capture distinguishing features more effectively, particularly in linguistically complicated situations. Second, we introduce the implicit stance reasoning module, to explore the potential stance tendencies in the text and their relationships with related topics. This method is effective for identifying content that misleads through stance shifting or emotional manipulation, because it can capture the implicit information behind the text. Finally, we integrate these two algorithms together to form a new framework, CL-ISR, which leverages the discriminative power of contrastive learning and the interpretive depth of stance reasoning to significantly improve detection effect.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Human-AI Co-Creation: A Framework for Collaborative Design in Intelligent Systems
cs.HC 2025-07 reject novelty 3.0 of 10

A study of 24 designers reports lower cognitive load and higher ideation fluency with AI assistance, and a three-tier framework for human-AI co-creation is proposed.

Reference graph

Works this paper leans on

16 extracted references · 14 canonical work pages · cited by 1 Pith paper

[1]

Social Media Misinformation and the Prevention of Political Instability and Mass Atrocities,

K. Hook and E. Verdeja, “Social Media Misinformation and the Prevention of Political Instability and Mass Atrocities,” Stimson Center, 2022

work page 2022
[2]

Machine Learning Strategies for Fake News Detection,

S. M. Alshamrani, M. A. Alzain, and M. A. Alzain, “Machine Learning Strategies for Fake News Detection,” Information, vol. 16, no. 3, p. 189, 2025

work page 2025
[3]

The Limits of Machine Learning Models of Misinformation,

D. Stepanova and R. Ross, “The Limits of Machine Learning Models of Misinformation,” AI & Society, vol. 40, no. 2, pp. 123–135, 2025

work page 2025
[4]

A Survey on Contrastive Self -Supervised Learning,

A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, and F. Makedon, “A Survey on Contrastive Self -Supervised Learning,” arXiv preprint arXiv:2011.00362, Oct. 2020

arXiv 2011
[5]

Implicit Stance Detection with Hashtag Semantic Enrichment,

L. Deng, Z. Sun, G. Du, X. Fang, and B. Zhang, “Implicit Stance Detection with Hashtag Semantic Enrichment,” Mathematics, vol. 12, no. 11, p. 1663, 2024

work page 2024
[6]

ToCoAD: Two-Stage Contrastive Learning for Industrial Anomaly Detection

Y. Liang, Z. Hu, J. Huang, D. Di, A. Su, and L. Fan, “Two -Stage Contrastive Learning for Industrial Anomaly Detection,” arXiv preprint arXiv:2407.01312, July 2024

work page Pith review arXiv 2024
[7]

Liu, Y., Shen, X., Zhang, Y., Wang, Z., Tian, Y., Dai, J., & Cao, Y. (2024). A systematic review of machine learning approaches for detecting deceptive activities on social media: Methods, challenges, and biases . arXiv. https://arxiv.org/abs/2410.20293

work page Pith review arXiv 2024
[8]

A Comparative Study on Linguistic Feature Selection in Sentiment Polarity Classification

Z. Liu, “A Comparative Study on Linguistic Feature Selection in Sentiment Polarity Classification,” arXiv preprint arXiv:1311.0833, Nov. 2013

work page Pith review arXiv 2013

Show all 16 references

[9]

A Survey on Social Network’s Anomalous Behavior Detection,

L. Xing, S. Li, Q. Zhang, H. Wu, and H. Ma, “A Survey on Social Network’s Anomalous Behavior Detection,” Complex & Intelligent Systems, vol. 10, pp. 5917–5932, May 2024

2024
[10]

The Spread of Low -Credibility Content by Social Bots,

C. Shao, G. L. Ciampaglia, O. Varol, K. Yang, A. Flammini, and F. Menczer, “The Spread of Low -Credibility Content by Social Bots,” Nature Communications, vol. 9, no. 1, p. 4787, Nov. 2018

2018
[11]

(2024) Design of Privacy - Preserving Personalized Recommender System Based on Federated Learning

Wang Y., Gong C., Xu Q., and Zheng Y. (2024) Design of Privacy - Preserving Personalized Recommender System Based on Federated Learning. 2024 8th International Workshop on Materials Engineering and Computer Sciences (IWMECS 2024)

2024
[12]

Large Pre -trained Language Models Contain Human -like Biases of What is Right and Wrong to Do,

P. Schramowski, C. Turan, N. Andersen, C. A. Rothkopf, and K. Kersting, “Large Pre -trained Language Models Contain Human -like Biases of What is Right and Wrong to Do,” arXiv preprint arXiv:2103.11790, Mar. 2021

2021 arXiv
[13]

Domain Generalization for Text Classification with Memory-Based Supervised Contrastive Learning,

Q. Tan, R. He, L. Bing, and H. T. Ng, “Domain Generalization for Text Classification with Memory-Based Supervised Contrastive Learning,” in Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022), Gyeongju, Republic of K orea, Oct. 2022, pp....

2022
[14]

Implicit Stance Detection with Hashtag Semantic Enrichment,

L. Dong, Z. Su, X. Fu, B. Zhang, and G. Dai, “Implicit Stance Detection with Hashtag Semantic Enrichment,” Mathematics, vol. 12, no. 11, p. 1663, May 2024

2024
[15]

Lyu, S. (2024). The Application of Generative AI in Virtual Reality and Augmented Reality. Journal of Industrial Engineering and Applied Science, 2(6), 1-9

2024
[16]

Lyu, S. (2024). Machine Vision -Based Automatic Detection for Electromechanical Equipment. Journal of Computer Technology and Applied Mathematics, 1(4), 12-20

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Social Media Misinformation and the Prevention of Political Instability and Mass Atrocities,

K. Hook and E. Verdeja, “Social Media Misinformation and the Prevention of Political Instability and Mass Atrocities,” Stimson Center, 2022

work page 2022

[2] [2]

Machine Learning Strategies for Fake News Detection,

S. M. Alshamrani, M. A. Alzain, and M. A. Alzain, “Machine Learning Strategies for Fake News Detection,” Information, vol. 16, no. 3, p. 189, 2025

work page 2025

[3] [3]

The Limits of Machine Learning Models of Misinformation,

D. Stepanova and R. Ross, “The Limits of Machine Learning Models of Misinformation,” AI & Society, vol. 40, no. 2, pp. 123–135, 2025

work page 2025

[4] [4]

A Survey on Contrastive Self -Supervised Learning,

A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, and F. Makedon, “A Survey on Contrastive Self -Supervised Learning,” arXiv preprint arXiv:2011.00362, Oct. 2020

arXiv 2011

[5] [5]

Implicit Stance Detection with Hashtag Semantic Enrichment,

L. Deng, Z. Sun, G. Du, X. Fang, and B. Zhang, “Implicit Stance Detection with Hashtag Semantic Enrichment,” Mathematics, vol. 12, no. 11, p. 1663, 2024

work page 2024

[6] [6]

ToCoAD: Two-Stage Contrastive Learning for Industrial Anomaly Detection

Y. Liang, Z. Hu, J. Huang, D. Di, A. Su, and L. Fan, “Two -Stage Contrastive Learning for Industrial Anomaly Detection,” arXiv preprint arXiv:2407.01312, July 2024

work page Pith review arXiv 2024

[7] [7]

Liu, Y., Shen, X., Zhang, Y., Wang, Z., Tian, Y., Dai, J., & Cao, Y. (2024). A systematic review of machine learning approaches for detecting deceptive activities on social media: Methods, challenges, and biases . arXiv. https://arxiv.org/abs/2410.20293

work page Pith review arXiv 2024

[8] [8]

A Comparative Study on Linguistic Feature Selection in Sentiment Polarity Classification

Z. Liu, “A Comparative Study on Linguistic Feature Selection in Sentiment Polarity Classification,” arXiv preprint arXiv:1311.0833, Nov. 2013

work page Pith review arXiv 2013