基于对比学习的深度聚类基线方法
A Deep Clustering Baseline Method Based on Contrastive Learning
宋鑫晶1
作者信息
- 1. 中国银联股份有限公司,上海 201201
- 折叠
摘要
聚类作为一种重要的无监督学习工具,在数据挖掘、图像处理等领域发挥着关键作用.研究提出了一种新的基于深度聚类的模型架构设计,包含对比实例生成网络、实例级对比学习网络和簇级对比学习网络.这种设计通过数据扩充和特征映射优化,实现了更有效的特征提取和聚类.在4个图像数据集上的实验表明,新模型在聚类准确率、归一化互信息和调整兰德指数等指标上表现优异,特别是在高维数据集上展现出卓越的性能.研究还证实了模型在不同改进方案下的灵活性和可用性,显示出其在深度聚类领域的前沿地位.
Abstract
Clustering,a vital unsupervised learning tool,plays a key role in fields such as data mining and image processing.A novel architecture for deep clustering is proposed,comprising contrastive instance generation networks,instance-level contrastive learning networks,and cluster-level contrastive learning networks.This design optimizes feature extraction and clustering through data augmentation and feature mapping.Experiments on four image datasets demonstrate the model's superior performance in clustering accuracy,normalized mutual information,and adjusted Rand index,especially in high-dimensional datasets.The research further confirms the model's flexibility and applicability under various improvement schemes,establishing its leading position in the field of deep clustering.
关键词
深度聚类/对比学习/深度学习Key words
deep clustering/contrastive learning/deep learning引用本文复制引用
出版年
2024
NETL
NSTL
万方数据