基于特征融合的部分有序深度森林模型  

Partial Ordered Deep Forest Model Based on Feature Fusion

作　　者：许行[1] 温萧轲 王文剑[1,2] XU Hang;WEN Xiaoke;WANG Wenjian(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;Key Laboratory of Computational Intelligence and Chinese Information Processing(Shanxi University),Ministry of Education,Taiyuan 030006,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]计算智能与中文信息处理教育部重点实验室(山西大学),太原030006

出　　处：《计算机工程与应用》2025年第7期165-175,共11页Computer Engineering and Applications

基　　金：国家自然科学基金(62206161,U21A20513,62076154,62276161);山西省重点研发计划项目(202202020101003,202302010101007);山西省基础研究计划项目(202303021221055);山西省高等学校科技创新项目(2020L0026);山西省“1331工程”项目。

摘　　要：部分有序数据是同时包含有序特征与无序特征的一类数据,其广泛存在于现实生活中。传统的有序分类方法或者将所有特征都视为有序特征,或者对有序与无序特征分别进行处理,忽略了二者之间的关系,这些方法难以有效解决部分有序数据上的分类问题。针对该问题,提出一种基于特征融合的部分有序深度森林模型,称为FFDF(feature fusion-based deep forest)。利用典型相关分析的思想,设计特征融合的贡献度计算方法,将有序特征和无序特征融合到同一特征空间,统一度量二者之间的关系。对融合的特征空间进行数据粒化,降低模型处理连续变量时的复杂性。设计融合空间下的特征矩阵输入级联森林,构建部分有序的深度森林模型。在来自UCI和WEKA的13个公共数据集上与部分单调决策树、有序分类模型、深度森林模型等六种方法进行比较实验,结果表明所提方法在准确性和平均绝对误差方面均优于对比方法;与集成模型深度森林gcForest和DF21进行了时间性能上的对比实验,结果表明所提方法在时间性能上优于对比方法。Partially ordered data is a class of data that contains both ordered and unordered features,and it exists widely in real life.Traditional ordered classification methods either regard all features as ordered features,or treat ordered and unordered features separately,ignoring the relationship between the two,these methods are difficult to effectively solve the classification problem on partially ordered data.To address the problem,a partially ordered deep forest model based on feature fusion,called FFDF(feature fusion-based deep forest),is proposed.Firstly,the idea of typical correlation analysis is utilized to design the contribution calculation method of feature fusion,which fuses ordered and unordered features into the same feature space,which helps to unify the measure of the relationship between the two.Then,the fused feature space is granulated,which can reduce the complexity of the model when dealing with continuous variable.Finally,the feature matrix on the fusion space is designed to input the cascade forest to construct a partially ordered deep forest model.Comparison experiments with six methods including partial monotone decision tree,ordered classification models REMT,OSDL,OLM,deep forest model gcForest and DF21 on 13 public datasets from UCI and WEKA,show that the proposed method outperforms the comparison methods in terms of accuracy and mean absolute error.In addition,the comparison experiments on time performance are conducted with the ensemble models gcForest and DF21,and the results show that the proposed method outperforms the comparison methods in terms of running efficiency.

关 键 词：有序分类 部分有序数据 特征融合 深度森林 典型相关分析 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]