数据聚类中基于浓度噪音消除的可视化参数选择方法

可视化技术的发展极大地提高了传统数据挖掘技术的效率.通过结合人类识别模式的能力,计算机程序能够更有效的发现隐藏在数据中的规律和信息.作为聚类分析的重要步骤,噪音消除一直都是困绕数据挖掘研究者的问题,尤其对于不同领域的应用,由于噪音的模型和定义不同,单一的数据处理方法无法有效而准确地去除域相关的噪音.本文针对这一问题,提出了一个新型的可视化噪音处理方法CLEAN.CLEAN的独特之处在于它设计的噪音处理技术和提出的可视化方法有机地结合在一起.噪音处理算法为可视化模型生成所需数据,同时针对噪音处理算法选择可视化方法,从而达到提高整个数据处理系统性能的目的.这样不仅降低了噪音去除过程中主观因素的影响,还可以帮助数据挖掘程序去除领域相关的噪音.同时源数据的质量,算法参数的选择和不同噪音去除算法的精确性都可以在所使用的可视化模型中反映出来.实验表明CLEAN能够有效地帮助空间数据聚类算法在噪音环境下发现数据的自然聚类.

A Visual Approach to Parameter Selection of Density-Based Noise Removal for Effective Data Clustering

Traditional visual data mining relies on visualization techniques to disclose implicit information and relationship among data through utilizing human capability of pattern recognition. As an important step in data clustering, noise removal is a challenging topic as domain-specific noise is not well defined and cannot be removed by generic process of data cleaning. This paper addresses two conjugated and reciprocal issues in the use of visualization in noise removal? choosing appropriate visualization techniques based on data removing methods, and designing processing algorithms that suit visualization. The goal is a synthesis of visualization techniques and data mining methods to enhance the overall performance while reducing the subjective factor in visual mining procedure. A visual data cleaning approach called CLEAN is proposed to assist spatial data clustering in four important aspects: removal of domain-specific noise, visualization of data quality, selection of algorithm parameters, and measurement of noise removing methods on parameter sensitiveness. Experiments show that the visualization models in CLEAN do assist effective discovery of natural spatial clusters in a noisy environment.