Data Mining from Hybrid Data with Numerical Attributes and Graph Structures

本研究项目的目的是为从具有数值属性和图结构的混合数据中进行数据挖掘提供理论基础。由于HTML/XML文件被认为是树形结构的数据，因此从树形结构的数据中发现特征模式的方法是有用的。基于遗传编程，我们实现了一个从给定的树结构数据的正反例中发现特征树结构模式的系统。我们的树结构模式是标记树模式。虽然标签树模式中的变量是可以由任意树替换的结构化变量，但这些变量被认为是树中的特殊边。然后，我们自然地将遗传编程应用到我们的发现系统中，这是一种针对树形结构对象的遗传方法。从实验或观测获得的数值数据中推断实值函数是从数值数据中进行数据挖掘的一种基本学习方法。递归实数是我们可以在计算机上处理的实数。因此，我们研究了递归实值函数的可学习性，如递归实值函数的预测和有限预测。我们还给出了针对树或图结构数据的各种学习算法，包括用于提取词之间的结构特征的算法和针对新引入的图语言类的正数据的多项式时间归纳推理算法。

The purpose of this research project is to give theoretical foundations of data mining from hybrid data with numerical attributes and graph structures. Since HTML/XML files are considered to be tree structured data, methods for discovering characteristic patterns from tree structured data are useful. Based on Genetic Programming, we have implemented a discovery system for characteristic tree structured patterns from given positive and negative examples of tree structured data. Our tree structured patterns are tag tree patterns. Although variables in a tag tree pattern are structured variables which can be substituted by arbitrary trees, these variables are considered to be special edges in a tree. Then we have naturally applied Genetic Programming, which is a genetic method for tree structured objects, to implementing our discovery system. Inferring real-valued functions from numerical data obtained from experiments or observations is a basic learning method for data mining from numerical data. A recursive real is a real number which we can deal with on a computer. So we have investigated learnabilities of recursive real-valued functions such as prediction and finite prediction of recursive real-valued functions. Also we have given various learning algorithms for tree or graph structured data, including an algorithm for extracting structural features among words and polynomial time inductive inference algorithms from positive data for newly introduced classes of graph languages.