面向高维混合数据的信息粒化与特征选择研究

Feature selection is the important research content in the fields of data mining, knowledge discovery and machine learning. Most of existing feature selection models and algorithms based on granular computing are mainly for single-type data, which achieve a lot of valuable research results. However, current data have some characteristics of dimensionality, heterogeneity and dynamicity, such that there are some serious challenges for existing feature selection methods. Therefore, in the context of some imperative and key problems on feature selection for high-dimension hybrid data, this project will deeply study effective methods of information granulation, and give efficient feature selection models and algorithms for high-dimension hybrid data. The main researches in this project are shown as follows: the layered granulation methods for multi-typed structured data, semi-structured data and unstructured data are provided; the feature evaluation method based on linear fusion is established; the feature selection algorithms under the calculation strategy for a continuous reduction search space on the object set and feature set are designed; the incremental updating mechanisms and algorithms of feature subset selection are provided when hybrid data increases dynamically; the effective evaluation mechanism for feature selection is provided. All of these will have important theoretical and practical significance for feature selection on high-dimension hybrid data, and will provide new ideas and methods for data processing and analysis under the background of the era of big data.

特征选择是数据挖掘、知识发现和机器学习等领域的重要研究内容。目前粒计算的特征选择方法在针对单一特征类型数据方面，已取得许多有价值的研究成果。然而，当今数据往往呈现出高维、多源异构和动态增长性等特点，使得现有特征选择方法在处理上述数据时面临严峻挑战。因此，本项目拟针对高维混合数据下特征选择亟待解决的关键难题，深入研究粒计算下高维混合数据的信息粒化方法，提出有效的特征选择模型和高效算法。具体包括：提出特征集下结构化、半结构化和非结构化数据的多层次信息粒化方法；构造基于线性融合的特征度量方法；设计基于对象集和特征集渐进式并行压缩策略的特征选择算法；针对动态增长的高维混合数据，提出特征子集的增量式更新机制和算法；建立特征选择的有效评估机制。本项目研究对高维混合数据的特征选择具有重要的理论意义和应用价值，为大数据时代下的数据分析和挖掘提供新思路和方法。

本项目针对高维混合数据下特征选择亟待解决的关键难题，深入研究了粒计算下高维混合数据的信息粒化方法，提出了有效的特征选择模型和高效算法。具体研究成果包括：提出了特征集下结构化、半结构化和非结构化数据的多层次信息粒化方法；构造了基于线性融合的特征度量方法；设计了基于对象集和特征集渐进式并行压缩策略的特征选择算法；针对动态增长的高维混合数据，提出了特征子集的增量式更新机制和算法；建立了特征选择的有效评估机制。本项目对高维混合数据的特征选择研究将为大数据时代下的数据分析和挖掘提供新思路和方法，具有重要的理论意义和应用价值。研究成果以第一作者或通讯作者发表研究论文7篇，其他相关论文5篇，其中以第一作者或通讯作者发表SCI检索论文6篇（二区论文4篇），EI检索1篇，软件著作权1项，培养在读硕士生3名，顺利完成项目的研究任务和目标。