Data Recovery Technology Based on Subspace Clustering

Abstract

High-dimensional data usually exist asymptotically in low-dimensional space. In this study, we mainly use tensor t-product as a tool to propose new algorithms in data clustering and recovery and verify them on classical data sets. This study defines the “singular values” of tensors, adopts a weighting strategy for the singular values, and proposes a tensor-weighted kernel norm minimization robust principal component analysis method, which is used to restore low-probability low-rank third-order tensor data. Experiments on synthetic data show that in the recovery of strictly low-rank data, the tensor method and weighting strategy can also obtain more accurate recovery when the rank is relatively large, which improves the volume of the rank. The proposed method combines the two and reflects its superiority through the restoration of 500 images under a small probability noise level.