Abstract:
The goal of this paper is to establish a widely applicable method for exploiting the sparsity in generalized eigenvector estimation. We propose an ℓ 1 -penalized extensio...View moreMetadata
Abstract:
The goal of this paper is to establish a widely applicable method for exploiting the sparsity in generalized eigenvector estimation. We propose an ℓ
1
-penalized extension of the Adaptive normalized quasi-Newton algorithm (Nguyen and Yamada, 2013 . To enhance sparsity in the estimate of the generalized eigenvector, the proposed adaptive algorithm maximizes a certain non-convex criterion with ℓ
1
penalty. A convergence analysis is also given for the proposed algorithm with decaying weight. Numerical experiments show that the proposed algorithm improves the subspace tracking performance in the situation where the covariance matrix pencil has sparse principal generalized eigenvector and is effective for recent sparsity-aware eigenvector analysis, e.g., sparse PCA.
Published in: 2018 IEEE Statistical Signal Processing Workshop (SSP)
Date of Conference: 10-13 June 2018
Date Added to IEEE Xplore: 30 August 2018
ISBN Information: