Visualizing large-scale high-dimensional data via hierarchical embedding of KNN graphs

• DOI: 10.1016/j.visinf.2021.06.002
• Publication Date: 2021-06-26T10:12:44Z
• Authors: Haiyang Zhu; Minfeng Zhu; Yingchaojie Feng; Deng Cai; Yuanzhe Hu; Shilong Wu; Xiangyang Wu; Wei Chen
• Keywords: Speedup; Convexity; Graph Embedding

Visualizing intrinsic structures of high-dimensional data is an essential task in analysis. Over the past decades, a large number methods have been proposed. Among all solutions, one promising way for enabling effective visual exploration to construct k-nearest neighbor (KNN) graph and visualize low-dimensional space. Yet, state-of-the-art such as LargeVis still suffer from two main problems when applied large-scale data: (1) they may produce unappealing visualizations due non-convexity cost function; (2) visualizing KNN time-consuming. In this work, we propose novel visualization algorithm that leverages multi-level representation achieve high-quality layout employs cluster-based approximation scheme accelerate layout. Experiments on various datasets indicate our approach achieves speedup by factor five compared yields aesthetically pleasing results.