講演抄録/キーワード |
講演名 |
2014-11-18 15:00
[ポスター講演]Robust Estimation under Heavy Contamination using Unnormalized Models ○Takafumi Kanamori(Nagoya Univ.)・Hironori Fujisawa(ISM) IBISML2014-68 |
抄録 |
(和) |
(事前公開アブストラクト) In this paper, our concern is to develop a new approach for robust data analysis on the basis of scoring rules. We propose a simple method of estimating not only parameters in the statistical model but also the contamination ratio of outliers. For this purpose, we use scoring rules with extended statistical models called unnormalized models. Also, regression problems are considered. We study complex heterogeneous contamination wherein the contamination ratio of outliers in a dependent variable may depend on independent variables. We propose a simple method to obtain a robust regression estimator under heterogeneous contamination. |
(英) |
In data analysis, contamination caused by outliers is inevitable, and robust statistical methods are strongly demanded.
In this paper, our concern is to develop a new approach for robust data analysis on the basis of scoring rules.
The scoring rule is a discrepancy measure to assess the quality of probabilistic forecasts. We propose a simple method
of estimating not only parameters in the statistical model but also the contamination ratio of outliers. Estimating the
contamination ratio is important, since one can detect the outliers in training samples based on the estimated
contamination ratio. For this purpose, we use scoring rules with extended statistical models called unnormalized
models. Also, regression problems are considered. We study complex heterogeneous contamination wherein the
contamination ratio of outliers in a dependent variable may depend on independent variables. We propose a simple method
to obtain a robust regression estimator under heterogeneous contamination. In addition, our method provides an
estimator of the expected contamination ratio that is available to detect the outliers in training samples. Numerical
experiments demonstrate the effectiveness of our method compared to conventional estimators. |
キーワード |
(和) |
/ / / / / / / |
(英) |
Scoring Rules / Unnormalized Models / Contamination Ratio / Regression / Heterogeneous Contamination / / / |
文献情報 |
信学技報, vol. 114, no. 306, IBISML2014-68, pp. 251-258, 2014年11月. |
資料番号 |
IBISML2014-68 |
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 |
IBISML2014-68 |