CPGAN: Curve Clustering Architecture Based on Projected Latent Vector of Generative Adversarial Network

• Published in: IEEE access 2020, Vol.8, p.86765-86776
• Main Authors: Zheng, Aiyu, Cai, Jianghui, Yang, Haifeng, Zhao, Xujun
• Format: Article
• Language: English
• Subjects: anomalous detection; Architecture; Clustering; Convolution; curve data; Data models; Datasets; Gallium nitride; generative adversarial network; Generative adversarial networks; Generators; Image analysis; Machine learning; Random noise; Robustness; Task analysis
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2020.2992887

Although Generative Adversarial Network(GAN) has obtained remarkable achievements in the image analysis and generation, its exploration in GAN-based curve clustering is still limited. The latent space of curve data is often used for clustering. However, the distance geometry in the latent space does not reflect the inherent clusters. In this paper, we propose CPGAN(Curve Clustering Architecture based on Projected Latent Vector of Generative Adversarial Network) for the clustering of curve dataset. Firstly, a novel GAN network structure, which utilizes a projector {P} (composed of the transposed convolutional network) to reconstruct the latent space of curve data, is proposed. CPGAN utilizes the concatenation of discrete code and Gaussian noise as a latent vector to preserve the implicit signal and structure of the cluster. Secondly, the loss function with two regularizations for CPGAN is proposed to guarantee the robustness and effectiveness of the model. Based on these, the jointly trained projector {P} is used to participate in the clustering process, while the generator can be used in the generating process. Finally, the spectral dataset from the LAMOST survey, the UCI dataset, and the UCR dataset are used as experimental data to evaluate clustering performance, the robustness of CPGAN, and further application on anomalous detection. CPGAN presents higher results than other methods.