基于类权重和最小化预测熵的测试时集成方法  

A test-time ensemble method based on class weights and prediction entropy minimization

作　　者：宋辉 张轶哲 张功萱[1] 孟元 SONG Hui;ZHANG Yizhe;ZHANG Gongxuan;MENG Yuan(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,Jiangsu,China)

机构地区：[1]南京理工大学计算机科学与工程学院,江苏南京210094

出　　处：《山东大学学报（工学版）》2024年第3期36-43,共8页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(62272232,62201263)。

摘　　要：针对传统集成学习方法忽略不同样本需使用不同模型权重的问题,提出一种基于类权重和最小化预测熵(class and entropy weights,CEW)的测试时集成方法。类权重为模型预测结果与验证集上各类概率对错分布的相似度,利用欧氏距离计算相识度;在最小化熵过程中,线性组合模型预测经过类权重模块加权后的输出,寻找最小预测熵对应的线性组合作为熵权重,提高集成模型预测能力。试验结果表明:在4个公开医学图像数据集上,CEW方法与最优单一模型相比,平均召回率提高0.23%~2.81%,准确率提高0.5%~2.54%;与DS方法相比,CEW方法平均召回率最多提高1.25%,准确率最多提高1.1%。基于CEW的测试时集成方法能够在测试时(无标签情况下)动态调整模型权重,比同类方法的预测精度更高。To address the issue of traditional ensemble learning methods overlooking the necessity for different model weights for varied samples,a test-time ensemble approach based on class and entropy weights(CEW)was proposed.Class weights were determined by the similarity between the model's predictive results and the distribution of correct and incorrect probabilities for each class on the validation set,calculated using Euclidean distance.During the entropy minimization process,the output from the linear combination of model predictions was weighted by the class weight module.The linear combination corresponding to the minimum predictive entropy was identified as the entropy weight,enhancing the predictive capability of the ensemble model.Experimental results showed that on four public medical image datasets,compared to the optimal single model,the CEW method improved the average recall rate by 0.23% to 2.81%,and accuracy by 0.5% to 2.54%.Compared to the DS method,the CEW method improved the average recall rate by up to 1.25% and accuracy by up to 1.1%.The test-time ensemble method based on CEW proved capable of dynamically adjusting model weights during testing(in an unlabeled situation),achieving higher prediction accuracy than similar methods.

关 键 词：测试时集成方法 医学图像分类 类权重 最小化熵 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]