( )8()1

()2016
26-22
ROBUST LOESS

/ /

2016/5/26 2016/4/17 : 2016/2/9 :

robust-loess loess
robust-loess
bisquare . loess Last Absolute Residuals

(
)1 :
) :

i yi yi

i 1, 2,...n

xi yi
x
. yi
Loess

robust
box ][2 ][1
) (2004
plot . ][4
William ) (2015 ] William(1979)[3
) and Devlin(1988 ]: [1][2
)4 Loess
( yi g ( xi ) i i 1, 2,...n
)
Loess ][3
)Cleveland(1979 ) g ( xi i
Lowess locally weight scatter plot
smoothing . yi

Lowess Losee
. )2 :
Robust Loess

. .
)3 : Residuals
. Robust-Loess
( )

22
 ( )8()1
()2016

: Loess
:
i yi yi
)a
:

3
: bisquar
x xi
wi (1 )
i )d ( x
( (1 )2 )2 i MAD
MAD
wi
0 i MAD x xi
x )d ( x
i


.
MAD
:

) MAD median(


.
) ( i )b
lowess
loess
)c
MAD

) ( i
MAD .
.
1-4 :Robust Loess
Losee


. robust
robust ( Last )bisquare Loess
Absolut Residual ) (LAR bisquare
) (LAR (
) .
bisquare :

23
 ( )8()1
()2016

2 = - 0.4992


Goodness of fit:
.
SSE = 0.9101
)5 :
R- Square = 0.8772
MATLAB 2015
Adjusted R-Square = 0.8747

i xi
RMSE=0.09637

( )1 .

.

( SSE
R2 )
SSE

R-Square


. 1 0

SSE ( yi yi )2

(y
) y
2
( : )1 loess
R 2
i

( yi y )2

robust-loess by bisquar
:

0 = 0.9887 x
1 = 1.138 0.5 : 0.2.3

2 = - 0.7756
g ( xi ) 0 1 x 2 x 2
Goodness of fit:
Loess
SSE = 0.9687
R- Square = 0.5908 0 =-0.03475
Adjusted R-Square = 0.5824

1 = 1.36
RMSE=0.09942

24
 ( )8()1
()2016

( )3 . ( )2 .

( : )3 robust-loess by ( : )2 robust-loess by
)(LAR bisquare

)6 : ) robust-loess by (LAR
:
loess -1

.
o = 1.028 , 1 = 0.9745
-2
loess
2 = - 0. 6433


Goodness of fit:
.
robust . -3 SSE = 0.9504

R- Square = 0.5985

Adjusted R-Square = 0.5903

RMSE=0.09848

25
 )1() 8(
)2016(

Robust locally weighted "William S.Cleveland -3 : ) 7

Journal of regression and smoothing robust-loess
"American statistical association
(1979) issue368 vol.74 spline robust
William S. Cleveland and sudan J. Devlin -4 .) (
Locally weighted regression: Approach "
Journal of American "analysis by local :

pp596- NO.403 vol.83 "statistical association " -1

.(1988) 610 "
. version 6 release 12"Learn Matlab" Matlab -5 .)2004(
" -2
"
.)2015(

The Technique of Robust Loess in Regression Analysis

ZAINB HASSAN RADHY

Al Qadisiya University

College of Computer Science and Information Technology

Abstract :
In this paper the robust-loess method is used to estimate the nonparametric regression function. The Loess is non-
robust method and used in case of outliers where it bases on the less squares in regression which affects by
presence of outliers. In this paper, the robust-loess has been implemented through applying the last absolute
residuals and bi-square techniques which enhanced robustness to the weighted least squares in loess.

26