深度超球面变分聚类网络  

Deep Hyperspherical Variational Clustering Network

作　　者：余南骏 吴昊旻 陈飞宇 刘学文 YU Nanjun;WU Haomin;CHEN Feiyu;LIU Xuewen(School of Mathematical Sciences,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学数学科学学院,重庆401331

出　　处：《重庆师范大学学报（自然科学版）》2024年第3期100-112,共13页Journal of Chongqing Normal University:Natural Science

基　　金：国家自然科学基金青年科学基金项目(No.12101098);重庆市自然科学基金面上项目(No.cstc2021jcyj-msxmX1053);重庆市教育委员会科学技术研究计划重点项目(No.KJQN201800116)。

摘　　要：基于变分自动编码器的深度聚类方法因较好的生成能力和聚类性能而受到广泛关注。然而现有的基于变分自动编码器的聚类算法通常需处理具有挑战性的证据下界(Evidence lower bound, ELBO)来实现聚类。同时,这些算法需要先验知识来假设类别分布,大多数方法还依赖于高斯分布或高斯混合模型。为了克服这些问题,提出了一种基于超球面变分自动编码器的端到端深度聚类网络。该方法采用2个后验分布的KL散度代替了复杂的ELBO,并将von Mises-Fisher混合分布作为潜在空间嵌入的模型。此外它最大化了潜在表示和预测聚类分配之间的扩展互信息以实现更有区分性和平衡性的分配。通过与最先进的深度聚类技术在基准数据集上进行比较,验证了所提出的深度聚类方法的有效性。所建立的模型为复杂数据特征学习和聚类分析提供了一套新的方法,拓展了贝叶斯推断方法在深度聚类上的应用。In recent years,the deep clustering approach based on variational autoencoders(VAE)has garnered significant attention due to its remarkable generative capabilities and clustering performance.However,existing VAE-based clustering algorithms typically revolve around dealing with the challenging Evidence Lower Bound(ELBO)to achieve clustering.Moreover,they often require prior knowledge of the class distribution,with most relying on Gaussian distributions or Gaussian mixture models.A deep clustering network based on Hyperspherical Variational Auto-Encoders(HVAE)is proposed,named DHVC(Deep Hyperspherical Variational Clustering Network).This approach substitutes the complex Evidence Lower Bound(ELBO)with the Kullback-Leibler Divergence of two posterior distributions and employs the von Mises-Fisher Mixture Model(vMFMM)distribution as an embedding in the latent space.Additionally,it maximizes the Extended Mutual Information(EMI)between latent representations and predicted cluster assignments for more discriminative and balanced allocations.The effectiveness of the proposed deep clustering method is validated through comparisons with state-of-the-art deep clustering techniques on benchmark datasets.The established model provides a novel approach for complex data feature learning and cluster analysis,extending the application of Bayesian inference methods in deep clustering.

关 键 词：KL散度 证据下界 超球面变分自动编码器 von Mises-Fisher混合模型 扩展互信息 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]