Conditional Generative Adversarial Network Based on Self-Attention Mechanism and VAE Algorithm and Its Applications

Generative Adversarial Networks (GANs) still face issues such as a lack of diversity in generated samples, incomplete encoding techniques, and a simplistic evaluation system. Based on this, the paper proposes a "Conditional Generative Adversarial Network based on self-attention mechanism and Va...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Traitement du signal Jg. 42; H. 1; S. 423 - 431
Hauptverfasser: Yang, Jianing, Zhao, Yanming, Jia, Zhiwei
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Edmonton International Information and Engineering Technology Association (IIETA) 01.02.2025
Schlagworte:
Adaptive sampling
Algorithms
Coding
Counterfeit
Credit card fraud
Data analysis
Datasets
Decision making
Fraud prevention
Generative adversarial networks
Higher education
Higher education institutions
Identification
Innovations
Methods
Neural networks
Optimization
Portfolio management
Portfolio performance
Stochastic models
Synthetic data
Time series
ISSN:0765-0019, 1958-5608
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Generative Adversarial Networks (GANs) still face issues such as a lack of diversity in generated samples, incomplete encoding techniques, and a simplistic evaluation system. Based on this, the paper proposes a "Conditional Generative Adversarial Network based on self-attention mechanism and Variational Autoencoder (VAE) Algorithm and Its Applications." The proposed algorithm consists of three sub-algorithms. The Variational Autoencoder (VAE) algorithm based on a self-attention mechanism adaptively constructs a latent space based on training data, thereby enhancing the diversity of generated samples. The A self-adaptive encoding method integrating self-attention and conditional vector projection. This method combines the self-attention mechanism and projection encoding algorithm to capture long-range dependencies in the data, addressing the issue of incomplete encoding techniques. Multi-metric Weighted Evaluation Algorithm is developed, which comprehensively evaluates the quality and diversity of generated samples, the conditional dependencies of the model, and the similarity between the distributions of input and generated samples. The evaluation metrics can be controlled adaptively through weight $\lambda_i$. The study constructs a financial dataset of higher education institutions containing 1,200 records and trains the proposed conditional GAN on this dataset. The network is then used to generate synthetic data for the detection of counterfeit data. Experimental results demonstrate that the proposed algorithm is feasible, stable, and shows comparative advantages.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0765-0019
1958-5608
DOI:10.18280/ts.420136