为了获得具有良好的分类准确性同时兼具可解释性的模型,提出了一种基于特征传递与重构的深度模糊分类器(deep fuzzy classifier based on feature transform and reconstruction,FR-DFC).基于深度学习的层次堆叠思想,数个用于特征传递的模糊系统(fuzzy systems for feature transform,FT_FS)和一个多原型模糊分类系统(multi-prototype fuzzy classification system,MPRFD_FS)堆叠在一起实现了模型的分类过程.其中,数个堆叠的FT_FS将数据从原始数据空间传递到高层特征空间中来探索数据中隐含的特征信息.MPRFD_FS则依据多个描述高层特征空间中类的分布特征的原型来实现分类.另一方面,利用多个用于特征重构的模糊系统(fuzzy systems for feature reconstruction,RE_FS)建立高层特征空间与原始数据空间的映射关系,并在原始数据空间中建立了一个便于理解的近似模糊分类器来保证FR-DFC具有可解释性.采用基于梯度下降和端到端的学习方式来优化模型的参数.优化的目标函数包含了分类损失函数和重构损失函数,能够同时保证模型的分类准确性和可解释性.实验结果表明:该分类器在提升分类准确性的同时兼具可解释性.
A Deep Fuzzy Classifier Based on Feature Transform and Reconstruction
To obtain a classifier with good classification accuracy and interpretability,a deep fuzzy classifier based on feature transform and reconstruction(FR-DFC)is proposed.In FR-DFC,several fuzzy systems(FT_FS)for feature transform and a multi-prototype fuzzy classification system(MPRFD_FS)are stacked together to realize the classification process of the model,based on the hierarchically stacked thought originated from deep learning.Specifically,the stacked FTFSs explore the hidden features in the data by transferring data from the original data space to the high-level feature space.MPRFD_FS,on the other hand,implements classification based on multiple prototypes that characterize the distribution of classifications in the high-level feature space.In addition,the proposed FR-DFC uses several fuzzy systems(RE_FS)for feature reconstruction to establish the mapping relationship between the high-level feature space and the original data space and establishes an understandably approximate fuzzy classifier in the original data space to ensure the interpretability of FR-DFC.Besides,FR-DFC utilizes gradient descent-based and end-to-end learning patterns to optimize the parameters of the model.The optimized objective function contains a classification loss function and a reconstruction loss function,which ensures both classification accuracy and interpretability of the model.Experimental results demonstrate that FR-DFC not only improves the classification accuracy but also possesses interpretability.
NETL
NSTL
万方数据
