摘要
深度规则分类器在处理大规模或高复杂度的分类任务时,模糊规则库的数量会过于庞大,导致其内部结构变得复杂,可解释性较差,测试时间也随着模糊规则数量的增加而线性递增,且在移动端设备内存有限的条件下无法保证内存的有效性.针对这些问题,提出一个新的内存有效的快速双层深度规则分类器,该分类器在深度规则分类器的基础上增设数据汇总和高描述性原型提取两个模块.仅当原型数量达到内存上限时执行一次数据汇总模块来删除部分原型,解决有限条件下内存不足无法训练出有效分类器的问题.高描述性原型提取模块将原型划分为底层和顶层的双层结构,底层由所有原型组成,用于展示分类器内部结构的全貌,顶层由少量高描述性原型组成,用于分类决策阶段.这样能有效地防止深度规则分类器在处理大规模数据时,因其复杂的内部结构导致较差的可解释性,同时提高分类决策阶段的效率.在基准数据集上的仿真实验验证了该方法的可行性和有效性.内存有效的快速双层深度规则分类器在处理大规模数据或数据流问题时,在达到较高分类性能的同时,保证模型的透明性、解释性和内存的有效性.
When deep rule-based classifiers deal with large-scale or complex classification tasks,the number of fuzzy rule bases will be too large resulting in complex internal structure,poor interpretability and linear in-crease of test time with the increase of the number of fuzzy rules.And the effectiveness of memory cannot be guaranteed under the condition of limited memory of mobile devices.To solve these problems,a new memory efficient fast double-layer deep rule-based classifier is proposed.Based on the deep rule-based classifier,two modules are added to the classifier,which are data summary and highly descriptive prototype extraction.Only when the number of prototypes reaches the upper limit of memory,execute the data summary module once to delete some prototypes,so as to solve the problem of limited memory that unable to train an effective classifiers.The highly descriptive prototype extraction module divides the prototype into two layers:the bottom layer and the top layer.The bottom layer is composed of all prototypes to show the whole picture of the in-ternal structure of the classifier.The top layer consists of a small number of highly descriptive prototypes for the classification decision-making stage,which effectively prevents the poor interpretability of deep rule-based classifier due to its complex internal structure when dealing with large-scale data,and improves the efficiency of classification decision-making stage.Simulation experiments on benchmark datasets verify the feasibility and effectiveness of this method.When dealing with large-scale data or data stream problems,the memory effective fast double-layer deep rule-based classifier achieves high classification performance and ensures the transparency,interpretability and memory effectiveness of the model.
作者
吕佳
肖锋
Lu Jia;Xiao Feng(College of Computer and Information Sciences,Chongqing Normal University,Chongqing,401331,China;National Center for Applied Mathematics in Chongqing,Chongqing Normal University,Chongqing,401331,China)
出处
《南京大学学报（自然科学版）》
CAS
CSCD
北大核心
2023年第3期446-459,共14页
Journal of Nanjing University（Natural Science）
基金
国家自然科学基金重大项目(11991024)
重庆市教委“成渝地区双城经济圈建设”科技创新项目(KJCX2020024)
重庆市高校创新研究群体(CXQT20015)。
关键词
规则分类器
模糊规则
内存有效
可解释性
原型
rule-based classifier
fuzzy rule
memory availability
interpretability
prototype
