| 講演抄録/キーワード |
| 講演名 |
2014-11-18 15:00
[ポスター講演]A note on least angle regression in orthogonal case ○Katsuyuki Hagiwara(Mie Univ.) IBISML2014-61 |
| 抄録 |
(和) |
本研究では,デザイン行列が直交性を満たす場合のLARS(least angle regression) を考える.これを便宜上,LARSO(LARS in tthe Orthogonal case) と呼ぶ.LARSO は,LARS のアルゴリズム的側面を理解する上で重要であるだけでなく,ノンパラメトリック回帰の問題に適用することを考えれば,wavelet denoising を例として,応用上も重要な役割を果たす.本研究では,LARSO が縮小推定を伴う貪欲法であることを示すとともに,k ステップにおける閾値が最小二乗水定量の絶対値のk + 1 番目に大きい値で与えられるsoft-thresholding 法となっていること示した.また,LARSO のCp タイプのモデル選択規準の簡単な導出法を示した.さらに,ノンパラメトリック直交回帰の問題において,このLARSO+Cp による方法とuniversal soft-thresholding やSUREshrink との関係を明確にするとともに,wavelet denoising の問題において,これらを数値的に比較した. |
| (英) |
LARS(least angle regression) is one of the sparse modeling methods. This paper considered LARS in the case of orthogonal design matrix, which we refer to as LARSO(LARS in the Orthogonal case). LARSO is not only an example
for understanding LARS but also is important in applications especially in the context of non-parametric regression including wavelet denoising. In this paper, we showed that LARSO is represented by a simple non-iterative algorithm. Interestingly, the resulting estimators of coefficients are shrinkage estimators under a greedy procedure. Based on this result, we found that
LARSO is exactly equivalent to a soft-thresholding method in which a threshold level at the kth step is the (k + 1)th largest absolute value of the least squares estimators. We also gave a simple proof of deriving a Cp type model selection criterion for LARSO. It is interpreted as a criterion not only for choosing the number of steps/coefficients of LARSO but also for determining an optimal threshold level in LARSO-oriented soft-thresholding method. Furthermore, in the context of non-parametric regression, we clarified relationship between LARSO and several methods such as the universal thresholding and SUREshrink in wavelet denoising. Throughout numerical experiments of application to wavelet denoising, we showed that LARSO with Cp type criterion outperforms the universal soft-thresholding method in terms of a generalization performance. |
| キーワード |
(和) |
LARS / 直交回帰 / soft-thresholding / wavelet denoising / / / / |
| (英) |
LARS / 直交回帰 / soft-thresholding / wavelet denoising / / / / |
| 文献情報 |
信学技報, vol. 114, no. 306, IBISML2014-61, pp. 199-206, 2014年11月. |
| 資料番号 |
IBISML2014-61 |
| 発行日 |
2014-11-10 (IBISML) |
| ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
| PDFダウンロード |
IBISML2014-61 |