Multi-View Unsupervised Feature Selection With Soft Label Learning and Tensor Low Rank Approximation

Advancements in information and storage technologies have led to the proliferation of high-dimensional multi-view data, necessitating robust feature selection methods. However, existing approaches often disregard data fuzziness and employ simplistic multi-view fusion strategies, thereby failing to simultaneously account for view diversity and consistency. To address these limitations, we introduce an unsupervised multi-view feature selection method, MESA, which integrates soft label learning and tensor low-rank approximation. Specifically, we first leverage the Fuzzy C-Means algorithm to construct an initial soft label matrix by measuring distances between data points and cluster prototypes. Next, we form a third-order tensor from the soft label matrices across multiple views and impose a tensor nuclear norm constraint to capture both view consistency and diversity. To achieve a unified framework for soft label learning and feature selection, we employ a sparse regression model. Additionally, we develop an efficient optimisation algorithm based on the alternating direction method of multipliers for iterative variable updates. Extensive experiments validate the effectiveness of our proposed approach, demonstrating notable performance improvements.