Abstract
An increasing amount of collected data are high-dimensional multi-way arrays(tensors), and it is crucial for efficient learning algorithms to exploit thistensorial structure as much as possible. The ever-present curse ofdimensionality for high dimensional data and the loss of structure whenvectorizing the data motivates the use of tailored low-rank tensorclassification methods. In the presence of small amounts of training data,kernel methods offer an attractive choice as they provide the possibility for anonlinear decision boundary. We develop the Tensor Train Multi-way Multi-levelKernel (TT-MMK), which combines the simplicity of the Canonical Polyadicdecomposition, the classification power of the Dual Structure-preservingSupport Vector Machine, and the reliability of the Tensor Train (TT)approximation. We show by experiments that the TT-MMK method is usually morereliable computationally, less sensitive to tuning parameters, and gives higherprediction accuracy in the SVM classification when benchmarked against otherstate-of-the-art techniques.