Graph Neural Network-Driven Hierarchical Mining for Complex Imbalanced Data

Muqing Li; Quanchao Lu; Shiyu Dou; Xiaoxuan Sun; Yankaiqi Li; Yijiashun Qi

REVIEW 5 major objections 5 minor 24 references

Graph Neural Network-Driven Hierarchical Mining for Complex Imbalanced Data

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A graph neural network embedding converted to a transaction database and mined with FP-growth finds more frequent patterns with higher minority-class coverage than raw-feature or PCA-reduced mining on a credit-card fraud dataset.

desk verdict A plausible GNN+FP-growth idea, but missing thresholds and inconsistent tables make the headline gains unverifiable; not ready for review. read the letter →

arxiv 2502.03803 v1 pith:BZQHVWDN submitted 2025-02-06 cs.LG

Yijiashun Qi , Quanchao Lu , Shiyu Dou , Xiaoxuan Sun , Muqing Li , Yankaiqi Li This is my paper · ORCID

classification cs.LG

keywords imbalanceddatagraphneuralnetworkhierarchicalminingfrequentpatternFP-growthminorityclasshigh-dimensionalcreditcardfrauddetection

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

This paper argues that embedding a high-dimensional imbalanced dataset into a sample-similarity graph, learning node representations with a graph neural network, and then converting those embeddings into a transaction database for FP-growth yields richer and more reliable frequent patterns than mining the original features directly. On the Credit Card Fraud Detection Dataset it reports 120 patterns with 5.8% average support, 87.2% average confidence, and 92.5% minority-class coverage, versus 85 patterns, 3.4% support, 74.8% confidence, and 68.9% coverage for original-feature FP-growth. The claim matters because it offers a concrete path from deep graph representation learning to classical pattern mining, making minority-class signatures visible enough to be extracted as rules. A hierarchical weighting step that upweights minority samples during training is what keeps the learned patterns from being dominated by the majority class.

What carries the argument

The engine is a two-stage pipeline. Samples become vertices of a similarity graph with Gaussian-kernel edge weights $e_{ij} = \exp(-\|x_i-x_j\|^2/\sigma^2)$, and a graph convolution layer updates each node representation by aggregating neighbor information; a minority-class weight $w_i$ reinforces the loss on minority nodes, and a contrastive local loss is added to the global classification loss. In the second stage, the learned embeddings are converted into a transaction database and mined with FP-growth, the standard frequent-itemset algorithm. Support, confidence, and minority-class coverage then measure whether the mined patterns are meaningful; the embedding-to-transaction conversion is the step that turns representation learning into a pattern-mining problem.

What would settle it

Run the deep-graph and raw-feature pipelines on the same Credit Card Fraud Dataset with identical, pre-specified frequent-pattern cutoffs (minimum support, minimum confidence, and the binarization threshold); if the deep-graph pipeline no longer yields more patterns with higher minority coverage, the claimed advantage is an artifact of the cutoff choice.

Watch

Extended reading notes

Core claim

The paper's central claim is that GNN-embedded features, binarized into a transaction database and mined with FP-growth, expose minority-class patterns that raw high-dimensional features hide. Its Table 1 states the gain directly: 120 patterns at 92.5% minority coverage for the deep-graph pipeline, against 85 patterns at 68.9% for original-feature FP-growth and 95 patterns at 78.3% for PCA-reduced FP-growth. The paper also reports that a 128-dimensional embedding is the best setting, and that constructing the sample graph with mutual information rather than KNN or a complete graph gives the strongest mining results.

Load-bearing premise

The reported comparison assumes that all three pipelines used the same cutoff values for how frequent a pattern must be and the same rule for turning continuous embeddings into yes/no transaction entries; the paper does not state those cutoffs.

Editorial extensions

If this is right

If the reported comparison is correct, GNN-embedded FP-growth recovers roughly 40% more minority-class frequent patterns than raw-feature FP-growth and lifts minority-class coverage from 68.9% to 92.5% on the tested fraud data.
Embedding dimension is a real hyperparameter for this pipeline: 128 dimensions gives the best reported trade-off, while 256 dimensions slightly degrades pattern count, support, and coverage.
Graph construction drives results: mutual-information graphs outperform KNN and complete graphs on all four reported metrics, so edge choice is a primary design lever.
Minority-class weighting during training is what prevents the majority class from swamping the learned patterns, meaning the mining gain is attributed to the hierarchical loss, not just the graph convolution.

Reading between the lines

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

The embedding-to-transaction step decouples representation learning from pattern mining, so the same embeddings could feed association-rule, sequence, or subgroup-discovery miners without changing the GNN.
A natural next test is to extend the hierarchical weighting to streaming data so that minority coverage adapts to distribution drift; the paper names this direction but does not implement it.
The same pipeline should transfer to other high-dimensional imbalanced domains such as medical anomaly prediction and user behavior analysis, which the paper lists as planned scenarios but does not evaluate.

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

Net: this is a reasonable idea wrapped in an unreliable experimental report. The pipeline—GNN embeddings fed into FP-growth for minority-class pattern mining—appears not to be in the cited literature, and the hierarchical minority weighting in the graph loss is a sensible design choice. Using the real credit-card fraud data and comparing against both raw features and PCA is the right kind of control. The method is not crazy; the paper is just not ready.

The load-bearing problem is that every headline number depends on unstated FP-growth thresholds and an unspecified binarization of continuous embeddings. Without the min_sup, min_conf, and the discretization cutoff, the 120 vs 85 patterns and the 92.5% vs 68.9% coverage could be threshold artifacts rather than evidence about the method. There is also no code, no error bars, no significance test, and no sensitivity analysis over the thresholds. A reader cannot reproduce the tables.

There are also internal inconsistencies. Table 1's deep graph row reports 87.2% confidence and 92.5% coverage, while Table 3's mutual information graph row reports 87.5% and 91.8% for what is described as the same best configuration. The introduction promises precision/recall/F1 evaluation; the experiments never measure them. The method section lists several hyperparameters (Gaussian bandwidth, balancing lambda, minority weight beta, hidden dimensions, layers) but gives no values or selection procedure.

The citation pattern is heavy on a tight cluster of recent arXiv and conference papers, but I would not make that a primary complaint; the experimental gaps are enough on their own.

Who is this for? A researcher wanting to test whether GNN embeddings can improve FP-growth on imbalanced tabular data might mine the idea and run their own controlled comparison. But as submitted, the paper is not a reliable source for the reported improvement. I would not cite it. It does not deserve a serious referee in this form; desk reject, or return with a demand for code, complete threshold settings, and a fixed, consistent results table. If those arrive, the idea could support a solid short paper.

Referee Report

5 major / 5 minor

Summary. The paper proposes a hierarchical mining framework for high-dimensional imbalanced data. It constructs a Gaussian-kernel similarity graph over samples, applies a graph neural network to obtain embeddings, converts those embeddings into a transaction database, and runs FP-growth to mine minority-class feature patterns. The reported experiments on the Credit Card Fraud Detection Dataset compare this pipeline with raw-feature FP-growth and PCA-reduced FP-growth, claiming higher pattern count, average support, average confidence, and minority-class coverage. The conclusion positions the method as a general framework for high-dimensional imbalanced-data mining.

Significance. If the empirical claims were supported, the idea of using GNN embeddings as the input to frequent-pattern mining for imbalanced high-dimensional data would be a plausible and potentially useful direction, particularly for financial fraud detection. The paper has the merit of identifying a concrete pipeline and testing it on a standard benchmark dataset. However, as presented, the central empirical claim is not backed by a reproducible experimental protocol: key mining thresholds are omitted, the comparison conditions are not controlled, there are no error bars or statistical tests, and the tables contain internal inconsistencies. The methodological novelty is also difficult to assess because the loss functions and graph construction are not fully specified. The paper offers no code, which further limits verification.

major comments (5)

[IV.B, Table 1] The comparison is uncontrolled because the FP-growth pipeline is underspecified. The paper never states the minimum-support threshold, the minimum-confidence threshold, or the binarization/discretization rule used to convert continuous GNN embeddings into a transaction database. These parameters are load-bearing: lowering min_sup can increase the number of frequent patterns and change average support, and the discretization rule determines which feature combinations can appear in a transaction. Without these values or a per-condition threshold grid, the reported gains of 120 vs 85 patterns, 5.8% vs 3.4% support, and 92.5% vs 68.9% coverage cannot be attributed to the proposed method rather than to threshold artifacts.
[Tables 1 and 3] The results in Tables 1 and 3 are internally inconsistent. Table 1 reports the Deep Graph Model Embedding + FP-Growth condition as 120 patterns, 5.8% average support, 87.2% average confidence, and 92.5% coverage, whereas Table 3 reports the Mutual Information Graph condition as 120 patterns, 5.8% support, 87.5% confidence, and 91.8% coverage. The paper never states which graph-construction method was used for Table 1, so the reader cannot know whether the headline result corresponds to the mutual-information graph or to a different configuration. Additionally, the text says four graph construction methods are evaluated, but Table 3 lists only three rows: the adaptive-threshold graph result is missing.
[IV.B, Tables 2-3] No error bars, confidence intervals, or statistical tests are reported for any of the experimental tables. The GNN involves stochastic training, and the minority class is only 0.172% of the data, so the observed differences (e.g., 120 vs 85 patterns, or the 128 vs 256 dimension comparison in Table 2) could easily lie within run-to-run variation. The claim of 'significantly outperforms' in the Abstract and Conclusion requires repeated runs and a significance test or at least variance reporting.
[I and IV] The Introduction's final paragraph promises that the method is evaluated with 'precision, recall, F1 value and other indicators,' but Section IV reports only pattern-mining metrics (number of patterns, average support, average confidence, minority-class coverage). No classification or predictive accuracy results are given for the proposed method or for any baseline, so the abstract's statements about 'efficiency and accuracy' and 'generalization ability' are not supported by the reported experiments.
[III] The method section is too underspecified to be reproducible. The local loss L_local is never defined; only a general mention of contrastive learning is given. The Gaussian bandwidth sigma, the minority weight beta, the balancing hyperparameter lambda, the number of GNN layers, the hidden dimensions, and the training procedure are all unspecified. Furthermore, the graph-construction alternatives compared in Table 3 (KNN, complete, mutual information, adaptive threshold) are introduced only there, with no formal definitions in Section III. The choice of the 128-dimensional embedding and the mutual-information graph in Tables 2-3 is made using the same evaluation metrics on the same dataset, with no held-out validation, which creates a selection-overfitting risk.

minor comments (5)

[Tables 2-3] The column header 'Number of modes' should be 'Number of patterns'.
[III] The minority-class weight definition contains a garbled symbol 'orityDmin'; the intended subscript appears to be 'D_minority', and the two cases should be typeset clearly.
[IV.B] Figure 2 is described as showing blue majority nodes and red minority nodes, but the figure itself is not reproduced in enough detail in the text to verify this; a higher-resolution figure with a caption explaining the colors would help.
[I] The Introduction uses 'depth map model' in one sentence; this should be 'depth graph model' for consistency with the rest of the paper.
[References] The reference list contains many arXiv preprints and 'unpublished' items; the authors should verify that all cited works are publicly available and are the most appropriate sources for the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's claims rest on empirical comparisons, not on definitions, fitted parameters, or self-citations that reduce to its own inputs.

full rationale

The paper does not contain a derivation chain in which a predicted quantity is defined in terms of the fitted quantity or vice versa. The central comparison is empirical: GNN embeddings are converted into a transaction database and mined with FP-growth, and the resulting pattern count, support, confidence, and minority coverage are compared against the same FP-growth pipeline applied to original and PCA-reduced features (Table 1). None of the reported metrics is constructed from the model parameters by definition, and no equation in Section III implies the pattern-mining outputs. The undisclosed FP-growth thresholds and binarization rules are a reproducibility and validity concern, not circularity: there is no quoted evidence that thresholds were chosen to force the reported numbers. The selection of embedding dimension (Table 2) and graph construction method (Table 3) using the same evaluation metrics on the same dataset is model selection or overfitting, not circular reasoning, because the chosen configurations are measured, not derived from the target metrics. References are background citations and do not carry a load-bearing uniqueness theorem or an ansatz that defines the method's output. Therefore no specific circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. The main unstated inputs are the many hyperparameters and thresholds that control the graph construction and the pattern-mining step, all of which are load-bearing for the reported numbers.

free parameters (6)

Gaussian kernel bandwidth sigma
Appears in the edge weight definition in Section III but the value is never reported.
Loss balancing weight lambda
Used in the optimization objective L = L_global + lambda * L_local in Section III, but never specified.
Minority class weight beta
Introduced to reduce majority-class weight contribution in Section III, but the actual value is not given.
GNN hidden dimensions and number of layers = 128 embedding dimension selected; other layer details missing
Table 2 tests 32, 64, 128, 256 dimensions and selects 128, but layer counts and widths are not stated.
FP-growth minimum support threshold
Directly controls the number of patterns and average support, yet no threshold is reported in Section IV.
FP-growth confidence threshold and embedding binarization cutoff
The conversion of continuous embeddings into a transaction database requires a binarization rule; neither that rule nor the confidence threshold is stated.

assumptions (4)

domain assumption Gaussian kernel similarities define a meaningful graph structure for the data.
Section III constructs edges with e_ij = exp(-||x_i - x_j||^2 / sigma^2); if the similarity graph is uninformative, the GNN embeddings inherit its noise.
domain assumption Frequent patterns mined from GNN embeddings correspond to real minority-class feature combinations.
Section IV.B treats FP-growth patterns as 'key feature combinations' without any semantic or external validation of the mined patterns.
domain assumption The mutual information graph construction is correctly defined and beneficial.
Table 3 selects the mutual information graph as the best method, but its construction is not described anywhere in the paper.
standard math GNN message passing captures global interdependencies among samples.
Section III relies on the standard graph convolution update to aggregate neighbor information, which is a known technique but an unproven assumption for this dataset.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Graph Neural Network-Driven Hierarchical Mining for Complex Imbalanced Data." pith.science (2026). https://pith.science/paper/BZQHVWDN

@misc{pith2026250203803,
  author       = {Pith},
  title        = {Pith review of: Graph Neural Network-Driven Hierarchical Mining for Complex Imbalanced Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZQHVWDN}},
  note         = {Machine review of arXiv:2502.03803}
}

read the original abstract

This study presents a hierarchical mining framework for high-dimensional imbalanced data, leveraging a depth graph model to address the inherent performance limitations of conventional approaches in handling complex, high-dimensional data distributions with imbalanced sample representations. By constructing a structured graph representation of the dataset and integrating graph neural network (GNN) embeddings, the proposed method effectively captures global interdependencies among samples. Furthermore, a hierarchical strategy is employed to enhance the characterization and extraction of minority class feature patterns, thereby facilitating precise and robust imbalanced data mining. Empirical evaluations across multiple experimental scenarios validate the efficacy of the proposed approach, demonstrating substantial improvements over traditional methods in key performance metrics, including pattern discovery count, average support, and minority class coverage. Notably, the method exhibits superior capabilities in minority-class feature extraction and pattern correlation analysis. These findings underscore the potential of depth graph models, in conjunction with hierarchical mining strategies, to significantly enhance the efficiency and accuracy of imbalanced data analysis. This research contributes a novel computational framework for high-dimensional complex data processing and lays the foundation for future extensions to dynamically evolving imbalanced data and multi-modal data applications, thereby expanding the applicability of advanced data mining methodologies to more intricate analytical domains.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 17 canonical work pages

[1]

Credit Default Prediction with Machine Learning: A Comparative Study and Interpretability Insights,

Y. Wang, Z. Xu, K. Ma, Y. Chen and J. Liu, "Credit Default Prediction with Machine Learning: A Comparative Study and Interpretability Insights," unpublished

work page
[2]

Dynamic Risk Control and Asset Allocation Using Q-Learning in Financial Markets,

M. Jiang, Z. Xu and Z. Lin, "Dynamic Risk Control and Asset Allocation Using Q-Learning in Financial Markets," Transactions on Computational and Scientific Methods, vol. 4, no. 12, 2024

work page 2024
[3]

Artificial Intelligence- Driven Risk Assessment and Control in Financial Derivatives: Exploring Deep Learning and Ensemble Models,

G. Huang, Z. Xu, Z. Lin, X. Guo and M. Jiang, "Artificial Intelligence- Driven Risk Assessment and Control in Financial Derivatives: Exploring Deep Learning and Ensemble Models," Transactions on Computational and Scientific Methods, vol. 4, no. 12, 2024

work page 2024
[4]

Object Detection for Medical Image Analysis: Insights from the RT-DETR Model,

W He, Y. Zhang, T. Xu, T. An, Y. Liang and B. Zhang, “Object Detection for Medical Image Analysis: Insights from the RT-DETR Model,” arXiv preprint, 2025

work page 2025
[5]

Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks

J. Wei, Y. Liu, X. Huang, X. Zhang, W. Liu and X. Yan, "Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks", 2024 5th International Conference on Machine Learning and Computer Application (ICMLCA), pp. 272-276, 2024

2024
[6]

Learning Self-Growth Maps for Fast and Accurate Imbalanced Streaming Data Clustering

Y. Xu, Y. Lee, R. Zou, et al., “LSROM: Learning Self-Refined Organizing Map for Fast Imbalanced Streaming Data Clustering,” arXiv preprint arXiv:2404.09243, 2024

work page Pith review arXiv 2024
[7]

Intrusion detection system using statistical query tree with hierarchical clustering approach,

P. V. N. Rajeswari and M. Shashi, “Intrusion detection system using statistical query tree with hierarchical clustering approach,” International Journal of Data Mining, Modelling and Management, 2024

work page 2024
[8]

Exploring Beyond Logits: Hierarchical Dynamic Labeling Based on Embeddings for Semi-Supervised Classification

Y. Ma, L. Jiao, F. Liu, et al., “Exploring Beyond Logits: Hierarchical Dynamic Labeling Based on Embeddings for Semi-Supervised Classification,” arXiv preprint arXiv:2404.17173, 2024

work page Pith review arXiv 2024

Show all 24 references

[9]

Data mining techniques for effective detection of distributed denial-of-service attacks,

Y. H. Lee, “Data mining techniques for effective detection of distributed denial-of-service attacks,” UTAR, 2024

2024
[10]

An Interpretable Bearing Fault Diagnosis Model Based on Hierarchical Belief Rule Base,

B. Zhao, Y. Qu, M. Mu, et al., “An Interpretable Bearing Fault Diagnosis Model Based on Hierarchical Belief Rule Base,” KSII Transactions on Internet & Information Systems, vol. 18, no. 5, 2024

2024
[11]

Leveraging Generative Adversarial Networks for Addressing Data Imbalance in Financial Market Supervision,

M. Jiang, Y. Liang, S. Han, K. Ma, Y. Chen, and Z. Xu, “Leveraging Generative Adversarial Networks for Addressing Data Imbalance in Financial Market Supervision,” arXiv preprint arXiv:2412.15222, 2024

2024 arXiv
[12]

Few-Shot Learning with Adaptive Weight Masking in Conditional GANs,

J. Hu, Z. Qi, J. Wei, J. Chen, R. Bao, and X. Qiu, “Few-Shot Learning with Adaptive Weight Masking in Conditional GANs,” in Proc. 2024 Int. Conf. Electron. Devices, Comput. Sci. (ICEDCS), pp. 435-439, Sep. 2024

2024
[13]

Robust Graph Neural Networks for Stability Analysis in Dynamic Networks,

X. Zhang, Z. Xu, Y. Liu, M. Sun, T. Zhou, and W. Sun, “Robust Graph Neural Networks for Stability Analysis in Dynamic Networks,” in Proc. 2024 3rd Int. Conf. Cloud Comput., Big Data Appl., Softw. Eng. (CBASE), pp. 806-811, Oct. 2024

2024
[14]

The Synergistic Role of Deep Learning and Neural Architecture Search in Advancing Artificial Intelligence,

X. Yan, J. Du, L. Wang, Y. Liang, J. Hu, and B. Wang, “The Synergistic Role of Deep Learning and Neural Architecture Search in Advancing Artificial Intelligence,” in Proc. 2024 Int. Conf. Electron. Devices, Comput. Sci. (ICEDCS), pp. 452-456, Sep. 2024

2024
[15]

Multi-Level Attention and Contrastive Learning for Enhanced Text Classification with an Optimized Transformer,

J. Gao, G. Liu, B. Zhu, S. Zhou, H. Zheng, and X. Liao, “Multi-Level Attention and Contrastive Learning for Enhanced Text Classification with an Optimized Transformer,” arXiv preprint arXiv:2501.13467, 2025

2025 arXiv
[16]

Leveraging Convolutional Neural Network-Transformer Synergy for Predictive Modeling in Risk- Based Applications,

Y. Wang, Z. Xu, Y. Yao, J. Liu, and J. Lin, “Leveraging Convolutional Neural Network-Transformer Synergy for Predictive Modeling in Risk- Based Applications,” arXiv preprint arXiv:2412.18222, 2024

2024 arXiv
[17]

Adaptive Transaction Sequence Neural Network for Enhanced Money Laundering Detection,

S. Long, D. Yi, M. Jiang, M. Liu, G. Huang, and J. Du, “Adaptive Transaction Sequence Neural Network for Enhanced Money Laundering Detection,” in Proc. 2024 Int. Conf. Electron. Devices, Comput. Sci. (ICEDCS), pp. 447-451, Sep. 2024

2024
[18]

Adaptive User Interface Generation Through Reinforcement Learning: A Data-Driven Approach to Personalization and Optimization,

Q. Sun, Y. Xue, and Z. Song, “Adaptive User Interface Generation Through Reinforcement Learning: A Data-Driven Approach to Personalization and Optimization,” arXiv preprint arXiv:2412.16837, 2024

2024 arXiv
[19]

Self-Supervised Credit Scoring with Masked Autoencoders: Addressing Data Gaps and Noise Robustly,

Y. Yao, “Self-Supervised Credit Scoring with Masked Autoencoders: Addressing Data Gaps and Noise Robustly,” J. Comput. Technol. Softw., vol. 3, no. 8, 2024

2024
[20]

Transforming Multidimensional Time Series into Interpretable Event Sequences for Advanced Data Mining

X. Yan, Y. Jiang, W. Liu, D. Yi, and J. Wei, "Transforming Multidimensional Time Series into Interpretable Event Sequences for Advanced Data Mining", arXiv preprint, arXiv:2409.14327, 2024

2024 arXiv
[21]

Calibration Learning for Few-shot Novel Product Description,

Z. Liu, M. Wu, B. Peng, Y. Liu, Q. Peng and C. Zou, "Calibration Learning for Few-shot Novel Product Description," Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 1864-1868, July 2023

2023
[22]

Computer Vision- Driven Gesture Recognition: Toward Natural and Intuitive Human- Computer Interaction,

F. Shao, T. Zhang, S. Gao, Q. Sun and L. Yang, "Computer Vision- Driven Gesture Recognition: Toward Natural and Intuitive Human- Computer Interaction," arXiv preprint arXiv:2412.18321, 2024

2024 arXiv
[23]

Dynamic User Interface Generation for Enhanced Human-Computer Interaction Using Variational Autoencoders,

R. Zhang, S. Wang, T. Xie, S. Duan and M. Chen, "Dynamic User Interface Generation for Enhanced Human-Computer Interaction Using Variational Autoencoders," arXiv preprint arXiv:2412.14521, 2024

2024 arXiv
[24]

Comparison of Norm-Based Feature Selection Methods on Biological Omics Data,

J. Song and Z. Liu, "Comparison of Norm-Based Feature Selection Methods on Biological Omics Data," Proceedings of the 5th International Conference on Advances in Image Processing, pp. 109 -112, November 2021

2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Credit Default Prediction with Machine Learning: A Comparative Study and Interpretability Insights,

Y. Wang, Z. Xu, K. Ma, Y. Chen and J. Liu, "Credit Default Prediction with Machine Learning: A Comparative Study and Interpretability Insights," unpublished

work page

[2] [2]

Dynamic Risk Control and Asset Allocation Using Q-Learning in Financial Markets,

M. Jiang, Z. Xu and Z. Lin, "Dynamic Risk Control and Asset Allocation Using Q-Learning in Financial Markets," Transactions on Computational and Scientific Methods, vol. 4, no. 12, 2024

work page 2024

[3] [3]

Artificial Intelligence- Driven Risk Assessment and Control in Financial Derivatives: Exploring Deep Learning and Ensemble Models,

G. Huang, Z. Xu, Z. Lin, X. Guo and M. Jiang, "Artificial Intelligence- Driven Risk Assessment and Control in Financial Derivatives: Exploring Deep Learning and Ensemble Models," Transactions on Computational and Scientific Methods, vol. 4, no. 12, 2024

work page 2024

[4] [4]

Object Detection for Medical Image Analysis: Insights from the RT-DETR Model,

W He, Y. Zhang, T. Xu, T. An, Y. Liang and B. Zhang, “Object Detection for Medical Image Analysis: Insights from the RT-DETR Model,” arXiv preprint, 2025

work page 2025

[5] [5]

Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks

J. Wei, Y. Liu, X. Huang, X. Zhang, W. Liu and X. Yan, "Self- Supervised Graph Neural Networks for Enhanced Feature Extraction in Heterogeneous Information Networks", 2024 5th International Conference on Machine Learning and Computer Application (ICMLCA), pp. 272-276, 2024

2024

[6] [6]

Learning Self-Growth Maps for Fast and Accurate Imbalanced Streaming Data Clustering

Y. Xu, Y. Lee, R. Zou, et al., “LSROM: Learning Self-Refined Organizing Map for Fast Imbalanced Streaming Data Clustering,” arXiv preprint arXiv:2404.09243, 2024

work page Pith review arXiv 2024

[7] [7]

Intrusion detection system using statistical query tree with hierarchical clustering approach,

P. V. N. Rajeswari and M. Shashi, “Intrusion detection system using statistical query tree with hierarchical clustering approach,” International Journal of Data Mining, Modelling and Management, 2024

work page 2024

[8] [8]

Exploring Beyond Logits: Hierarchical Dynamic Labeling Based on Embeddings for Semi-Supervised Classification

Y. Ma, L. Jiao, F. Liu, et al., “Exploring Beyond Logits: Hierarchical Dynamic Labeling Based on Embeddings for Semi-Supervised Classification,” arXiv preprint arXiv:2404.17173, 2024

work page Pith review arXiv 2024