[1]程玉胜,孙鸿飞,余钟萍.融合路径长度的多路径层次标签分类方法[J].计算机技术与发展,2025,(07):140-147.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0060]
CHENG Yu-sheng,SUN Hong-fei,YU Zhong-ping.Multi-path Hierarchical Labeling Classification by Fusing Path Lengths[J].,2025,(07):140-147.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0060]
点击复制
融合路径长度的多路径层次标签分类方法(
)
《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]
- 卷:
-
- 期数:
-
2025年07期
- 页码:
-
140-147
- 栏目:
-
人工智能
- 出版日期:
-
2025-07-10
文章信息/Info
- Title:
-
Multi-path Hierarchical Labeling Classification by Fusing Path Lengths
- 文章编号:
-
1673-629X(2025)07-0140-08
- 作者:
-
程玉胜1; 2; 孙鸿飞2; 余钟萍2
-
1. 安庆师范大学 智能感知与计算重点实验室,安徽 安庆 246133;
2. 安庆师范大学 计算机与信息学院,安徽 安庆 246133
- Author(s):
-
CHENG Yu-sheng1; 2; SUN Hong-fei2; YU Zhong-ping2
-
1. Key Laboratory of Intelligent Perception and Computing,Anqing Normal University,Anqing 246133,China;
2. School of Computer and Information,Anqing Normal University,Anqing 246133,China
-
- 关键词:
-
层次标签; 标签分类; 路径长度; 多路径选择; 自顶向下分类
- Keywords:
-
hierarchical label; label classification; path length; multi-path selection; top-down classification
- 分类号:
-
TP183
- DOI:
-
10.20165/j.cnki.ISSN1673-629X.2025.0060
- 摘要:
-
标签分层分类方法从根节点出发,选择不同路径逐步细化分类。 针对现有方法忽略了分类时选择不同长度路径会面临不同风险的问题,提出了一种融合路径长度的多路径层次标签分类方法(MCPL)。 首先,采用自顶向下的递归方法,通过逻辑回归获得到达不同节点路径的概率;其次,根据节点的位置信息计算不同节点间的路径长度,利用路径长度为路径赋权,使用赋权后的父节点路径概率和当前节点路径概率以更新当前节点的路径概率;最后,在不同层级,依照节点间的兄弟关系在每个层级选择多个可能的粒度类别,将最后选择的多个类别经过分类器进行再次分类。 在 DD、F194、Car196、VOC、CLEF 和 Bridges 数据集上进行实验,相较于六种分层分类方法中最好的结果,MCPL 的样本分类准确率指标平均提高了 2. 4% ,层次分类指标平均提高了 0. 36% ,层次结构诱导误差指标平均降低了 1. 4%。 实验结果表明,MCPL能够有效提高分类性能。
- Abstract:
-
The hierarchical labeling classification method starts from the root node and gradually refines the classification by selecting different paths. Aiming at the problem that the existing methods ignore the different risks faced when choosing paths of different lengths during classification,a Multi-path Hierarchical Labeling Classification method by Fusing Path Lengths (MCPL) is proposed. Firstly,a top-down recursive approach is employed to obtain the path probabilities to reach different nodes using logistic regression. Secondly,the path lengths between different nodes are calculated based on the positional information of the nodes,and the path length is used to assign weights to the paths. The weighted parent node path probabilities and current node path probabilities are then used to update the current node’s path probability. Finally,at different levels,multiple possible granularity categories are selected based on the sibling relationships between nodes,and the final selected categories undergo further classification by a classifier. Experiments on the DD,F194,Car196,VOC,CLEF,and Bridges datasets show that compared to the best results among six hierarchical classification methods,MCPL’s sample classification accuracy increases by an average of 2. 4% ,the hierarchical classification metric improves by an average of 0. 36% ,and the hierarchical structure-induced error metric decreases by an average of 1. 4% . Experimental results demonstrate that MCPL can effectively improve classification performance.
更新日期/Last Update:
2025-07-10