一種用于數(shù)據(jù)流分類的遞歸反向傳播算法

打開文本圖片集

中圖分類號(hào)：TP181；TP183 文獻(xiàn)標(biāo)志碼：A

A Recursive Back Propagation Algorithm for Data Stream Classification

LIU Zhanhua， WEN Yimin， LIU Xiang （School of Computer Science and Information Security & School of Software Engineering， GuilinUniversityofElectronic Technology，Guilin541OO4，Guangxi，China）

Abstract：To enhancethelearning abilityof deep neural network model，a recursive back propagation algorithmfordata streamclassfication was proposedto solvethe problemoflowclasification accuracydue toconcept driftinthe traditional deep neural network.Theproposedalgorithm combined the powerful data stream learning ability ofonlinegradient descent algorithm with the fast convergence characteristic of recursive least square method.When the concept drift occurred in thedata stream，the neural network model wastrained graduallbyusing recursive least square method，after reaching arelativelystable state，online gradient descent algorithm was switched to further trainthedeep neural network model，achieve deeperdata stream learning，andoptimize the clasification performanceoftedeep neural network model. The effctivenessof the proposedalgorithm wasverified insomeartificialdata setsandrealdatasets.Theresults show that the proposed algorithm hasexcelentadaptability toconcept drift，and theaccuracyof datastream clasification exceeds those of many algorithms thatonly use online gradient descent algorithm or recursiveleast square method to train neural network model.

Keywords：onlinedeep learning；online gradient descent algorithm；recursive least square method；back propagation; deep neural network；concept drift

近年來(lái)，深度學(xué)習(xí)在眾多應(yīng)用領(lǐng)域取得了顯著成就[1-3]，然而，深度神經(jīng)網(wǎng)絡(luò)模型（DNN）的學(xué)習(xí)面臨諸多問(wèn)題，包括梯度消失、特征重用率下降[4]鞍點(diǎn)和局部最小值問(wèn)題[5]、龐大的參數(shù)調(diào)整量、訓(xùn)練過(guò)程中內(nèi)部協(xié)變量偏移[6、正則化器選擇困難、超參數(shù)難以確定等。（剩余12776字）