Fast and Provable Algorithms for Sparse Principal Component Analysis
Salvato in:
| Pubblicato in: | PQDT - Global (2025) |
|---|---|
| Autore principale: | |
| Pubblicazione: |
ProQuest Dissertations & Theses
|
| Soggetti: | |
| Accesso online: | Citation/Abstract Full Text - PDF Full text outside of ProQuest |
| Tags: |
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| Abstract: | Principal component analysis (PCA) is a well-known statistical method for feature extraction and dimension reduction, widely used for data analysis. However, traditional PCA encounters problems of overfitting and loss of explainability in high-dimensional settings, particularly when the number of variables exceeds the sample size. Sparse PCA overcomes these limitations by introducing sparsity into the principal components, offering a robust alternative to PCA and obtaining more interpretable results. In this thesis, we investigate the spiked covariance model and the spiked Wigner model in sparse PCA.We first explore the spiked covariance model, which aims to recover a sparse unit vector from noisy samples. From an information-theoretic perspective, Ω(k log p) observations are sufficient to recover a k-sparse p-dimensional vector v. However, existing polynomial-time methods require at least Ω(k 2 ) samples for successful recovery, highlighting a significant gap in sample efficiency. To bridge this gap, we introduce a novel thresholding-based algorithm that requires only Ω(k log p) samples, provided the signal strength λ = Ω(∥v∥ −1 ∞ ). We also propose a two-stage nonconvex algorithm that further enhances estimation performance. This approach integrates our thresholding algorithm with the truncated power iteration, achieving the minimax optimal rate of statistical error under the desired sample complexity. Numerical experiments validate the superior performance of our algorithms in terms of estimation accuracy and computational efficiency.Secondly, we study the spiked Wigner model, which aims to recover a s-sparse d-dimensional unit vector u from a d×d noisy matrix. The information theoretical lower bound of the signal strength required to estimate u is β = Ω(√ s log d). In contrast, the signal strength required for existing polynomial-time methods is at least Ω( e s), leading to a notable gap. To close this gap, we propose a new thresholding-based algorithm that requires only Ω(√ s log d) signal strength, given ∥u∥∞ = Ω(1). We also design a two-stage nonconvex method that further improves estimation accuracy. This approach combines our thresholding algorithm with the truncated power iteration, achieving the constant error in limited iterations under the desired signal strength. Empirical results show the advanced performance of our algorithms in terms of the estimation error and computational cost. |
|---|---|
| ISBN: | 9798263313760 |
| Fonte: | ProQuest Dissertations & Theses Global |