基于单向流模型的自适应张量链式学习算法  

Adaptive tensor train learning algorithm based on single-aspect streaming model

作　　者：马宝泽 李国军 邢隆[1,2] 叶昌荣 MA Baoze;LI Guojun;XING Long;YE Changrong(School of Optoelectronic Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Lab of Beyond LoS Reliable Information Transmission,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[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

出　　处：《通信学报》2023年第8期27-36,共10页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2019YFC1511300);国家自然科学基金资助项目(No.62201113,No.U22A2006);重庆市重点研发计划基金资助项目(No.cstc2021ycjh-bgzxm0072);重庆市教委科学技术研究基金资助项目(No.KJQN202300625)。

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。An adaptive tensor train(TT)learning algorithm for the online decomposition problem of high-order tensors in single-aspect streaming model was investigated.Firstly,it was deduced that single-aspect streaming increment only changes the dimension of temporal TT core.Secondly,the forgetting factor and regularization item were introduced to construct the objective function of exponentially weighted least-squares.Finally,the block-coordinate descent learning strategy was used to estimate the temporal and non-temporal TT core tensors respectively.Simulation results demonstrate that the proposed algorithm is validated in terms of increment size,TT-rank,noise and time-varying intensity,the average relative error and operation time are smaller than that of the comparison algorithms.The tensor slice reconstruction ability is superior than that of the comparison algorithms in the video adaptive analysis.

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 

分 类 号：TN911.6[电子电信—通信与信息系统]