A Tensor Based Big Data Efficicent Computation And Multimodal Analysis Apppoach

With the continual expansion of ubiquitous sensing devices and social media,various data are generated from different spaces(e.g.,cyber,physical,and social spaces)and form the cyber-physical-social big data with multisource,heterogeneous,and complex characteristics.Tensor-based big data representation model can represent and fuse various types of data derived from different fields.Tensor-based multidimensional correlation analysis method can also comprehensively analyze the global data and provide more accurate services for practical applications.However,there still exists some urgent challenges in the process of tensor-based big data analysis,such as repeated computation caused by streaming data,huge execution time caused by massive data scale,high energy consumption caused by complicated calculation,and poor analysis result caused by complex relationship.Therefore,this paper focuses on exploring a series of methods to implement the big data efficient computation and multimodal analysis by exploiting tensor train(TT)decomposition and high order singular value decomposition(HOSVD).The main contributions are as follows:(1)To avoid the repeated computation caused by streaming data,an incremental TT decomposition method is proposed.First.we propose some theorems to compute the TT result of zero-padding tensor based on the TT result of the original tensor.Then,we present an approach to compute the TT result of the updated tensor of the original and adding zeropadding tensors in TT format.Furthermore,the orthogonal and compressed method for the final TT cores is studied.The incremental TT decomposition method can avoid the repeated decomposition of the original tensor and improve the computational efficiency.It can provide a new method to compute the TT decomposition for the streaming data in the big data era.(2)To alleviate the problem of long execution time and explosive intermediate result during tensor computations,we propose a TT based computation method for most tensor operations and their parallel implementations.First,we construct a TT based big data processing framework including data representation,fusion,storage,computation,analysis,and application.Then,a series of TT based computation methods for most tensor operations are presented,they can be directly implemented based on decomposed TT cores and the result also maintains TT format.Furthermore,we put forward a parallel computation architecture including inter-core parallel and inner-core parallel models.The proposed TT based tensor computation methods are directly implemented in parallel based on the decomposed loworder TT cores.It provides a new computation theory and realization method to greatly improve the calculation efficiency of tensor computations.(3)To alleviate the problem of high energy consumption caused by the complex calculations,a set of tensor-based cloud-edge computing optimization methods are proposed.First,we construct a triple-plane cloud-edge computing architecture including edge devices,edge servers and cloud to collaboratively complete computational tasks.Then,some cloudedge computing optimization models are presented including total energy consumption,total execution time,system reliability,and users' quality of experience based on some tensorbased representation models.Furthermore,a multi-objective optimization and task assignment algorithm is proposed to reasonably allocate the computational tasks.The proposed cloud-edge computing optimization method not only takes full advantage of edge devices,edge servers and cloud,but also synergistically considers the influence of multiple objectives,it provides a general optimization framework for big data computation.(4)To improve the effect of big data analysis,a series of tensor-based multimodal analysis methods are proposed.In the prediction domain,we first propose a TT-based multivariate Markov transition method and an algorithm to calculate the stationary eigentensor of multivariate Markov and further realize multimodal predictions.The method is implemented in parallel only on some low-order TT cores,it can greatly reduce the execution time and save the memory overhead on the premise of guaranteeing the high prediction accuracy.Besides,we further propose a general multivariate multi-order Markov transition method based on tensor unified product,then present an iterative algorithm to calculate its stationary joint eigentensor to realize multimodal predictions.The proposed multimodal prediction method based on stationary joint eigentensor can significantly improve the prediction accuracy.In the recommendation and clustering domain,a set of incremental HOSVD based multi-dimensional correlative analysis methods are proposed.They can realize the personalized recommendation under different contexts and multi-modal clustering and further realize the adaptive construction of learning community under different situations.The multi-dimensional correlation analysis method has higher recommendation accuracy and clustering performance because it is analyzed based on the global data.These proposed tensor-based big data multimodal analysis methods can provide a new way to improve the effect of big data analysis.