符号数据的无监督学习:一种空间变换方法

Unsupervised Learning from Categorical Data:A Space Transformation Approach

近年来符号数据的无监督学习在模式识别、机器学习、数据挖掘和知识发现等诸多领域扮演着越来越重要的角色。然而现有的针对符号数据的聚类算法(经典的K-modes系列算法等),相比数值型数据的聚类算法,在性能方面仍然有很大的提升空间。其根本原因在于符号数据缺乏类似数值数据那样清晰的空间结构。为了能够有效地发掘符号数据内在的空间结构,采用了一种全新的数据表示方案:空间变换方法。该方法将符号数据映射到相应的由原来的属性组成的新的维度的欧氏空间中。在这一框架的基础上,为了找到符号数据更有代表性的模式,结合CarreiraPerpin提出的K-modes算法进行无监督学习。在9个常用的UCI符号数据集上进行了测试,与传统的符号数据聚类算法进行了实验比较,结果表明几乎在所有的数据集上提出的方法都是更加有效的。

The unsupervised learning method of categorical data plays a more and more important role in such areas as pattern recognition,machine learning,data mining and knowledge discovery in the recent years.Nevertheless,in view of many existing clustering algorithms for categorical data（the classical k-modes algorithm and so on）,there is still a large room for improving their clustering performance in comparison with the performance of clustering algorithms for numeric data.This may arise from the fact that categorical data lack a clear space structure as that of numeric data.To effectively discover the space structure inherent in a set of categorical objects,we adopted a novel data representation scheme：a space transformation approach,which maps a set of categorical objects into a corresponding Euclidean space with the new dimensions constructed by each of the original features.Based on the new general framework for categorical clustering,we employed the Carreira-Perpin＇s K-modes algorithm for clustering to find more representative modes.The performance of the new proposed method was tested on the nine frequently-used categorical data sets downloaded from the UCI.Comparisons with the traditional clustering algorithms for categorical data illustrate the effectiveness of the new method on almost all data sets.