Research on Nonnegative Tucker3 Decomposition in Krylov Sub-Space Calculation for Feature Extraction

The method of nonnegative Tucker3 decomposition usually has to face the problem of data overfitting in the iteration procedure, which leads to slow convergence and low accuracy. Thus the solution of conjugate lanczos method is researched to accelerate the iterative calculation or decrease the computing time. The novel algorithm takes advantage of implicit thick-restart lanczos method with crescent basis vectors for saving computer cost and improving the property of alternative products between vectors. Lastly, the proposed method not only extracts the feature from MIT_CBCL dataset but also achieves the required efficiency and accuracy through experiments.