Support vector machine—recursive feature elimination (SVM-RFE)

I

200 202 w 29‘

L RECEIVE TEST FORMULATE .____._2_____ OUTPUT FRO“ PROBLEM SVM ’ ’\, 2I8 COLLECT TRAINING POST-PROCESS DATA TEST OUTPUT /D 220 203 OPTIONALLY 222 F PREPROCESS TRAINING DATA N O OPTIMAL SOLUTION? Y ES 226 20 4 I 224 SELECT KERNEL FOR SVM 205 I ADJUST KERNEL COLLECT LIVE SELECTION DATA INPUT 208 PRE-PROCESSED J TRAINING D T O M, svtviA A T (‘N PRE-PROCESS LIVE DATA 22 8 I TRAIN SVM INPUT ( GENERATE OPTIMAL 230 REM" \/R 210 ‘ 1 DATA 2II OPTIONALLY 232 “CE; LIVE D POST-PROCESS D OUTPUT TRAINING OUTPUT POST-PROCESS LIVE COLLECT F TESTDATA f\' ____‘_)..1.m’UT 2 3 4 2| 2 I PRE—PROCESS ' END k,-A TESTD ATA 2 36 2I 4 INPUT PRE-PRO CESSED TEST DATA m-ro -T TRATNED SVM F I G _ 2

V 2I6

INITIALIZE BIN AND CLASS TRACKING f\" VARIABLES 304 CALCULATE DATA 306 __.___.___,,_ ENTROPY FOR EACH '\/ CANDIDATE BIN so 8 TEST SEQUENTIAL BIN COMBINATIONS OF CUTOFF ALL CUTS PLACED? DETERMINATION 310 TRY DIFFERENT NUMBER ADJUST NUMBER OF CUTS? 01-" CUTS OUTPUT 3 | l 3I2 CLASSIFICATION

3l4

END

FIG. 3

I‘ “ ' ' ' ' ' “ _ ' ' ‘ _ ' ' ‘ ' ' ' “ ' _ - ' - ' _ — ’ ' ' ” _ ‘ _ ’ ’ ' ” ' “ ' ' ‘ ' ' ' ' - “ “ - - ' - - ' ‘ _ _ - ' ’ ' ' ' ‘ - ' ' - - ' “ ' "“'|
I l
{SYSTEM MEMORY;'l5i=\;-[_f;i5§'fj'('3'é'S"S'if\'@'; 1075A }/7000
1 /0775 P RAM 1" I
'.- ---------------------- --=5 MODULE(S_) '=
:s OPTTMAL ;; ----- --; ---------------
:
; ;CATEGCR\ZAT\CN§ §P°SL{;§,%g§§A$'NG; / r 1023 ;
IE PROGRAM 5: MODULE(S) 1’ 1045 .
:§_____[\{'9RE{';‘§$?3)_M_:::::::::::::::::::::::::::::: sus / ; L%)‘(é%>.v8f;EA
. ................... 2 5 CONTROLLER -
OPERATING §VEC1's(¥R");gOR2-3RAM§ NETWORK
: 5 .... ‘N1-ERBACE I 1054
5 1065/ 1070/1020 5 1060
I 1024 : r
I ; REMOTE
5 ¢0N%'¥§LLER lNPUTlO\UTPUT BUS COMPUTER
E / 1040 /1044?
' DISK DRIVE‘ VlDEO SCSI PARALLEL SERIAL T
: CONTROLLER ADAPTER INTERFACE INTERFACE INTERFACE ; 1055
5 103° 1032 1042 \1041 \1043 E
' l===1 I ‘=—'_'=I==' I‘ |
1 ____________________ __‘_“_‘i'_____% ____________________ -1954 .............. _____!/057
1050 1053 I MODEM
\\ I H U WIDE AREA
H my 53 I FIG- 4 O 000 1052 1056 Mougg KEYBOARD 1038

9:10 17 was no: ‘0I ‘Wt luawd ‘Sn

ZH ‘E8V‘S60‘8 Sfl

1316 \ OPTIMAL OUTPUTC 1314A /__ 1314B OU'l'PUTC1 OUTPUTC2 1312A 1312B 2'” LEVEL 2'” LEVEL SUPPORT VECIOR SUPPORT VECTOR IVIACI-IJ‘N'E C1 MACHINEC2 (KERNEI. 1) 1310 (KERNEL 4) A A 2"’LEvEI. INPUT 1308A /‘ 1308B _ED'I'II\1.AI. OPTIMAL UTPLIT A T 4 OUTPUTB 1306A 3063 1306C 1306D \ /’ 1 -\ /T OUTPUT A1 OUTPUT A2 OU'IT>UT BI OUTPUTB2 1302A 1302B 1302C 13021) 1“LEvEI.SUPI>oIu' FLEVELSUPPQRT I“LEVI-:I. SUPPORT I“LEVEI. SUPPORT VECTQR MACHINE VECTOR MACHINE VECTOR MACHINE VECTORMACHINE A1 A2 B1 B2 (KERNEL 1) (KERNEL 2) (KERNEL 1) (KERNEL 3) ISTLEVEL 1S‘l‘LEV'.EL INPUTA \ INPUTS \ 130-IA 1304B

FIG. 5

TO~TEST SUCCESS RATE
-I- LEAVE-ONE-OUT QUALITY CRITERION
-0- (SMOOTHED) PREDICTOR OF OPT.
_0 4 ~ TEST SUCCESS RATE

' -A- EPSILON (THEORETICAL ERROR BAR)

'0'60{l3'»456789l0

1-0 ' ‘ ' ‘ l I l \ |
0.9 -

0.8
0.?

Success Rate

0 .6
0.5
0 .4"

-0-TEST SUCCESS RATE

-I- LEAVE~ONE-OUT QUALITY CRlTERlON

0.3 T -Q- (SMOOTHED) PREDICTOR OF OPT.
TEST SUCCESS RATE

-A‘ EPSILON (THEORETICAL ERROR BAR)

0.2

0
0 1 2 3 4 5 6 7 8 9 10
log2(numbcr of genes)