Integrating feature and graph learning with low-rank representation

We propose a new subspace clustering method that integrates feature and manifold learning while learning a low-rank representation of the data in a single model. This new model seeks a low-rank representation of the data using only the most relevant features in both linear and nonlinear spaces, which helps reveal more accurate data relationships in both linear and nonlinear spaces, because data relationships can be less afflicted by irrelevant features. Moreover, the graph Laplacian is updated according to the learning process, which essentially differs from existing nonlinear subspace clustering methods that require constructing a graph Laplacian as an independent preprocessing step. Thus the learning processes of features and manifold mutually enhance each other and lead to powerful data representations. Extensive experimental results confirm the effectiveness of the proposed method.