面向多视角证据信息融合的高效特征选择方法  

An Efficient Feature Selection Approach with Multi-View Evidence Information Fusion

作　　者：樊晓雪 杨光 鞠恒荣 丁卫平[1] 黄嘉爽 杨习贝[2] FAN Xiaoxue;YANG Guang;JU Hengrong;DING Weiping;HUANG Jiashuang;YANG Xibei(School of Artificial Intelligence and Computer Science,Nantong University,Nantong 226019,Jiangsu,China;School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212003,Jiangsu,China)

机构地区：[1]南通大学人工智能与计算机学院,江苏南通226019 [2]江苏科技大学计算机学院,江苏镇江212003

出　　处：《昆明理工大学学报(自然科学版)》2025年第1期72-84,157,共14页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：国家自然科学基金项目(62006128,U2433216,62102199,62076111);江苏省自然科学基金项目(BK20231337);江苏高校青蓝工程项目([2024]14);南京大学计算机软件新技术国家重点实验室资助项目(KFKT2024B30);江苏省高等学校自然科学研究面上项目(24KJB520032);江苏省双创博士计划项目((2020)30986);中国博士后科学基金项目(2022M711716);江苏省研究生科研与实践创新计划项目(SJCX24_2016).

摘　　要：特征选择能在复杂数据中选出有效特征,提高信息处理效率.但是现有的粒计算模型在选取信息粒时仅考虑距离度量,忽略了样本间的其他联系.为了解决这一问题,作者提出了一种基于两阶段多视角邻域证据熵的特征选择方法.首先,根据稀疏约束获得每个样本的自适应k值,并通过稀疏约束和距离度量进行融合形成多视角邻域信息粒,然后,对该信息粒中的样本进行检测,进一步删除弱相关样本,降低信息粒的不确定性.其次,将稀疏信息引入证据理论中,形成新的信任函数,并与邻域熵结合构造邻域证据熵,能有效地反映数据的确定性和不确定性信息.接着,利用邻域证据熵评估特征的重要性,以此实现特征的选择.最后,在9个公共数据集上进行实验验证,结果表明:本文提出的方法在粒度构建和分类精度方面都优于其他算法,能降低信息的不确定性.此外,本文方法也应用于精神分裂症的脑区选取,其结果能有效地提高对精神分裂症的预测能力.Feature selection can select effective features in complex data and improve the efficiency of information processing.However,the existing granular computing model only considers the distance measure when selecting information granularity,ignoring other connections between samples.To solve this problem,this paper proposes a feature selection method based on two-step multi-view neighborhood evidence entropy.First,the adaptive k-value of each sample is obtained according to the sparsity constraint.The multi-view neighborhood information granularity is formed by fusing the sparsity constraint and the distance measure.Then,the samples in this information granularity are detected,and weakly correlated samples are further removed to reduce the uncertainty of information granularity.Secondly,sparse information is introduced into the evidence theory to form a new credibility-based function.It can combine with neighborhood entropy to construct neighborhood evidence entropy,which can effectively reflect the certainty and uncertainty information of data.Then,the neighborhood evidence entropy is used to evaluate the importance of features as a measure to achieve feature selection.Finally,experimental validation is conducted on nine public datasets.The results show that the proposed method outperforms other algorithms in terms of granularity construction and classification accuracy,reducing the uncertainty of information.Additionally,the method is applied to the selection of brain regions in schizophrenia,and the results effectively enhance the predictive ability for schizophrenia.

关 键 词：粒计算 多视角 信息融合 邻域证据熵 特征选择 

分 类 号：R318[医药卫生—生物医学工程] TP18[医药卫生—基础医学]