树形子图的统一向量表示及其多类标分类关键技术研究

Graph (network) used for describing interactions between objects has many real-world applications. We can achieve high valuable information from these networks by using graph mining techniques. In many applications, the content features of each node in graph which are usually heterogeneous/multiple are represented by a tree structure (i.e., each node is regarded as a tree-structured subgraph). Unlike traditional graph classification, tree-structured subgraph classification has three challenging issues: (1) features of each subgraph are organized in a tree structure; (2) two subgraphs have multiple types of relations; and (3) subgraphs may have multi-labels. The essential issues of this project are finding how to design an appropriate model to represent the feature space of tree-structured subgraphs, studying effective methods for integrating heterogeneous features, and optimizing the classification algorithms under a multi-relational multi-label learning environment. In this project, on the basis of a hierarchical framework to represent the tree structure of each graph node, we aim at studying the representation techniques for tree-structured subgraphs and the multi-label classification techniques for multi-relational graphs. The contributions of this research are threefold: (1) a unified vectorial representation framework for tree-structured subgraph including a multi-layer subspace learning algorithm and a multi-layer dual wing harmonium model, which aims to solve the fusion problem of content features of each node; (2) an iterative multilabel classification algorithm for multi-relational graph, which integrates the multi-relational information into the multilabel classification learning process; (3) a manifold learning based semi-supervised multilabel classifier for multi-relational graph, which enables us to solve the small sample problem in real life applications.

图（网络）作为刻画物体之间相互作用的工具，具有广泛的应用范围。运用图挖掘技术，我们可以从中得到高价值的信息。在很多应用中，每个图节点所包含的内容信息往往是异构/多特征的，并以树形组织结构呈现（即树形子图）。树形子图分类具有与传统的图分类所不同的三个重要问题：树形结构、多关系和多类标。其中的关键科学问题是：如何在树形结构的特征空间表示模型下，研究高效的特征融合方法，从而在多关系多类标环境下优化树形子图的分类算法。课题在分析含树形子图的多关系图数据特征基础上，重点研究树形子图的特征表示及其多类标分类问题。创新在于：（1）树形子图的统一向量表示模型，包括多层子空间学习算法和多层双翼谐振算法，用于解决节点的内容特征融合问题；（2）基于迭代学习的多关系图多类标分类算法，用于融合多关系信息并解决多类标分类问题；（3）基于流形学习的多关系图多类标半监督分类算法，用于解决小样本环境下的学习问题。

图（网络）作为刻画物体之间相互作用的工具，具有广泛的应用范围。运用图挖掘技术，我们可以从中得到高价值的信息。在很多应用中，每个图节点所包含的内容信息往往是异构/多特征的，并以树形组织结构呈现（即树形子图）。树形子图分类具有与传统的图分类所不同的三个重要问题：树形结构、多关系和多类标。其中的关键科学问题是：如何在树形结构的特征空间表示模型下，研究高效的特征融合方法，从而在多关系多类标环境下优化树形子图的分类算法。课题在分析含树形子图的多关系图数据特征基础上，重点研究树形子图的特征表示及其多类标分类问题。研究内容包括：1）含树形子图的多关系图表示方法；2）基于多层子空间学习算法的树形子图统一向量表示模型；3）基于多层双翼谐振算法的树形子图统一向量表示模型；4）基于迭代学习的多类型多关系图多类标分类算法；5）基于流形学习的多关系图多类标半监督分类算法；6）多关系图的多类标分类实验平台。依托此项目，共发表14篇SCI期刊论文(包括9篇IEEE Trans.论文)和7篇EI论文，并申请了5项国家发明专利。