DRO框架下不平衡分類損失函數(shù)重加權(quán)優(yōu)化

打開文本圖片集

關(guān)鍵詞：不平衡分類；數(shù)據(jù)標(biāo)簽不確定性；加權(quán)標(biāo)簽分布穩(wěn)健損失

中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）08-024-2428-09

doi：10.19734/j. issn.1001-3695.2024.10.0480

Optimization of re-weighted loss function for imbalanced classification under DRO framework

Li Jiajing’，Lin Geng2+（1.Schoolofttics&istcs，jnalUiersitzo57in;2holopuer>Uversity，F(xiàn)uzhou ，China）

Abstract：The skewed distributionofclassesoftenleads clasification models toneglecttheimportanceof minorityclasses，favoringthe majorityones，whichcanrender models incapableof accurate classification in multi-classimbalancedtasks.Existingresearch focusesonthe studyof data balancing strategies andloss function tuning，ignoring the problemofuncertaintyin labelinginformation，werelabelsmaybewongornoisyTheuncertaintymakesclasifiercorrectclasificationmorechallenging.Thispaperproposedanewlossfunction，calledweightedlabeldistributionallyrobustKullback-Leibler，whichoptimisedthepredictivedistributionundertheworst-casescenario，toaddressthechangesanduncertaintiesindatadistributionfor theimbalanced clasification task．Basedonadistributionallyrobust framework，thisapproach merged prior informationand label weights tofocus onminorityclasss andadapt to labeluncertainty.Inaddition，this paper proposedasimulation method forimbalanceddatasets thatused MonteCarlosimulations toprovideamorecomprehensiveevaluationof theperformanceof eachlossfunctionunderdiferent classesandatdifferent levelsofquantitativevariance.Experimentalresultsonsimulated， UCI and Kagledatasetsshowthatthe proposed method performs wellwith imbalanced dataandachievesamoderate improvement in top-k accuracy， F1 -scores，precision and recall.

Key words：multi-classimbalanced task；data label uncertainty；weighted labeled distribution robust loss

0 引言

在現(xiàn)實中，數(shù)據(jù)集往往呈現(xiàn)出一種傾斜的不平衡狀態(tài)，即某些類別的樣本數(shù)量會遠(yuǎn)遠(yuǎn)多于其他類別。（剩余20929字）