一种基于模特征的增量式张量Tucker分解方法

An incremental tensor tucker decomposition method based on model characteristics

随着数据量的爆炸式增长,边缘计算在大数据处理中的作用愈加重要.现实应用中产生的数据通常建模表示成高阶增量式张量的形式,增量式张量Tucker分解是一种高效挖掘高阶海量数据中隐藏信息的方法.针对传统增量式张量分解忽视张量模特征对分解过程的影响、分解结果不能较好保留原始数据特征的问题,提出一种基于模特征的增量式张量Tucker分解方法 ITTDMC (incremental tensor tucker decomposition based on mode characteristics).首先,用模长增量决定增量因子矩阵更新顺序,以此降低更新顺序带来的重构误差;其次,根据模熵变化比决定增量因子矩阵更新权重,使分解结果更准确保留各模特征;然后,将过往时刻的模特征和更新参数记录在指导张量中,遇到模特征相似的增量数据时直接使用来指导张量中参数的更新,避免重复计算,降低时间开销;最后,在合成和真实数据集上进行大量的实验,实验结果表明ITTDMC在模特征明显的数据集上能显著降低(最高可达29%)增量式张量的重构误差.

With the explosive growth of data volume,edge computing plays an increasingly important role in big data processing.In general,the data generated by real applications is modelled and represented as high-order incremental tensors.Recently,the incremental tensor tucker decomposition is deemed as an efficient approach to extract the information inherent in those high-order massive data.As the traditional incremental tensor decomposition usually ignores the influence of tensor model characteristics on the decomposition process,it is rather difficult for the decomposition results to preserve the overall characteristics of the original data.To address these issues,we propose an incremental tensor tucker decomposition method ITTDMC(incremental tensor tucker decomposition based on mode characteristics)based on mode characteristics.First,the update order of the increment factor matrix is determined by the increase of mode length,for reducing the reconstruction error caused by the update order.Next,the update weight of the incremental factor matrix is computed according to the changing ratio of the mode entropy,such that the decomposition results enable to maintain the characteristics of each module more accurately.Furthermore,the previous model characteristics and update parameters are recorded in a guide tensor.When the incremental data with similar model characteristics needs to be processed,the corresponding update parameters of the guide tensor can be directly employed to reduce the computational costs.Extensive experiment results on the synthetic and real data sets exhibit that the ITTDMC can greatly reduce the reconstruction error of incremental tensor(up to 29%)for those data sets with strong model characteristics.