Parallel HOSVD With Its Incremental Computation

As a high order data structure, tensor has been widely used by many big-data applications. The tensor HOSVD becomes very popular because of its efficient way of exploring tensor data' intrinsic features. However the computational efficiency of HOSVD still is the major bottleneck.With regards to this challenge, there are two main problems to be solved urgently. Firstly, how to cut the high order big tensor data into many smaller blocks, and combine the result of each block into a whole, more convenient, rapid and accurate representation on distributed systems. Secondly, how to avoid the re-computation of HOSVD on the data as the incremental data keep coming. This thesis presents a novel tensor data distribution approach and an efficient parallel implementation of its HOSVD. Furthermore, the thesis proposes a new incremental HOSVD method developed from the Jacobi-based SVD method. Our experimental results demonstrate our parallel HOSVD with its incremental computation methods have a considerably improvements in term of both parallel speedup and efficiency.