摘要
SOM(self-organizing map)所具有的拓扑保持特性使之可用来对高维数据进行低维展现,但由于数据间的距离信息在映射到低维空间中固定有序的神经元上时被丢掉了,因此数据的结构通常是被扭曲了的.为了更自然地展现数据的结构,提出了一种新的基于SOM的数据可视化算法——DPSOM(distance-preserving SOM),它能够按照相应的距离信息对神经元的位置进行自适应调节,从而实现了对数据间距离信息的直观展现.特别地,该算法还能自动避免神经元的过度收缩问题,从而极大地提高了算法的可控性和数据可视化的质量.
Due to the topology-preserving nature, the SOM（self-organizing map）algorithm can be used to visualize the high-dimensional data. However, due to the fixed regular lattice of neurons, the distance information between the data is lost, and thus the structure of the data may often appear in a distorted form. In order for the map to visualize the structure of the data more naturally, the distance information or the similarity information between the data should be preserved as much as possible on the map directly through the positions of the neurons, along with the topology. To do this, the positions of the neurons should be adjustable on the map. In this paper, a novel position-adjustable SOM algorithm, i.e., DPSOM （distance-preserving SOM）, is proposed, which can adaptively adjust the positions of the neurons on the map according to the corresponding distances in the data space and thus can visualize the structure of the data naturally. What＇s more, the DPSOM algorithm can automatically avoid the excess contraction of the neurons without any additional parameter, thus greatly improving the controllability of the algorithm, and the quality of data visualization.
出处
《计算机研究与发展》
EI
CSCD
北大核心
2006年第3期429-435,共7页
Journal of Computer Research and Development
基金
国家自然科学基金项目(60443003)~~
作者简介
邵超，1977年生，博士研究生，主要研究方向为人工神经网络、机器学习、数据可视化等．（sc_flying@163．com）
黄厚宽，1940年生。教授，博士生导师。主要研究方向为人工智能、机器学习、数据仓库、数据挖掘、多智能体系统等．
