2023年8月JournalonCommunicationsAugust2023第44卷第8期通信学报Vol.44No.8基于单向流模型的自适应张量链式学习算法马宝泽1,2，李国军1,2，邢隆1,2，叶昌荣1,2（1.重庆邮电大学光电工程学院，重庆400065；2.重庆邮电大学超视距可信信息传输研究所，重庆400065）摘要：针对单向流模型中高阶张量在线分解问题，研究了一种自适应张量链式（TT）学习算法。首先，推导出单向流增量仅改变时序TT核的维度；然后，引入遗忘因子和正则项构造指数权重最小二乘目标函数；最后，利用块坐标下降学习策略分别估计时序和非时序TT核。对所提算法在增量大小、TT秩、噪声和时变强度等方面分别进行了验证，结果表明，所提算法的平均相对误差和运算时间均小于对比算法，并在视频自适应分析中表现出优于对比算法的张量切片重构能力。关键词：自适应学习算法；张量链式分解；单向流模型；泛在数据流中图分类号：TN911.6文献标志码：ADOI:10.11959/j.issn.1000−436x.2023154Adaptivetensortrainlearningalgorithmbasedonsingle-aspectstreamingmodelMABaoze1,2,LIGuojun1,2,XINGLong1,2,YEChangrong1,21.SchoolofOptoelectronicEngineering,ChongqingUniversityofPostsandTelecommunications,Chongqing400065,China2.LabofBeyondLoSReliableInformationTransmission,ChongqingUniversityofPostsandTelecommunications,Chongqing400065,ChinaAbstract:Anadaptivetensortrain(TT)learningalgorithmfortheonlinedecompositionproblemofhigh-ordertensorsinsingle-aspectstreamingmodelwasinvestigated.Firstly,itwasdeducedthatsingle-aspectstreamingincrementonlychangesthedimensionoftemporalTTcore.Secondly,theforgettingfactorandregularizationitemwereintroducedtoconstructtheobjectivefunctionofexponentiallyweightedleast-squares.Finally,theblock-coordinatedescentlearningstrategywasusedtoestimatethetemporalandnon-temporalTTcoretensorsrespectively.Simulationresultsdemonstratethattheproposedalgorithmisvalidatedintermsofincrementsize,TT-rank,noiseandtime-varyingintensity,theaveragerelativeerrorandoperationtimearesmallerthanthatofthecomparisonalgorithms.Thetensorslicereconstructionabili-tyissuperiorthanthatofthecomparisonalgorithmsinthevideoadaptiveanalysis.Keywords:adaptivelearningalgorithm,tensortraindecomposition,...
