BA7ES DECISIoN TtER Bayes1an classif er

tD elae prablem)
Lenstder a classifcaton task d two elAsses w, od

w2 m ich owe patem- belns

Ow am is to design a cassiher h a t clasag

An unknenpalen m the most prbeblu h e

t o c l a s e s w, m d wz2

Assume h a t the probabiLahus CwD and pew2) e

T h e prababluy densly fFanehim p(r/o:), C=,2

desenbe he distribuho eature vees ln
erch e he elases. t cAn be echmakd PemL

tha avAilable i n n g d A t
The pef pczAst) somebmes refennd to as

ksiihrod fiunchey wt win respeet b*.

hhe
T h e ondihomal prshabLy PCwr), represen
fhat e palem belsngs
beloas
unkmown
h e probab1k
P te
the respechve
elas
w , iven that
N e e t o r t a k s fhe vaue
correspendia
feature
semáwme ruened to as
pCw) is
a posteriosi probabihis

t Bayes Rmle
According

=
PCz%) PCui)
PLW)
PC PCi) =P
is the pdf S} *,
tre pC«)
pCx) PC/r) P) pe)-P
 rmle c o n
l a s s 1 f i c a h o n

The
he Byes

be s t as,
t & elassif
class1f
PCw)>PCua/z),
to
z is class1g-e
f p/z)< plw/),
t 2

aelen the palem

The case of 2qnaliy t
eher sf the hwd c a s e s
can be assigned to
deeisis Can e equivAlently be baze
The
en he ínegaities,

PCTA) pC)pC z/wa) PW2)

a priori pbssChies
are eqnal,
te
e - PCwe) = Tham aome eqnahm

becsmes,
PlT)2pCr/w2)
>Figuse b e l w shows 'x
mple a t
ewpnhudeela%es
prbede
Snd shows the vrisho

PC/ur ), i =
,2
ea snge feaure
Simple Case
pCrl

o
-R2
 2

The dated n e a t z s a Thretnsld pmihama

4me K2
the teawe sfpnce Mto tro g n s ,,
rulu f>r all
Accordiigo the Bayes d e eis1 on
h a claçs rdesidesw am
VAmes 4 2 in ,

he fgwe that the

1+ s obvns -frsm .
ndecd
unaloideble
are
erors
deeisLon
far an * lia in
finile prababiuly
h e r e is
same h i e belog
at u
Th R2 e a B and
em
deensm is 1
m Ass w . Then Snr

decisis eror
T h s he totn prsbebky s
ease 4 tMD eqipsbable casses.
The
3 3ven by,

Pe- Pa/4,) daz / s ) dz

tet <hsdid undur
me s egmal to area

e Curves
.show hat fhe Bayesian ckss1fierIS ophma
wrth res pect to minimzing the classifiaho
error probabitG

Mini mi21n he classi ticshay er pab1l

will show hat the BayesiaL
n his proof we

chsier is ophmal witl repectlo minimigme

heelars1kiahon e r e poaalM

Pro s
space
the eatwee spale-
he region
Let Ri b e md R2
ecide nin er w 2
m wich we
correpondmma regi
2 w2 Then
be nhe

em errer s made E3,, hngh t

belongs
Wa er
i ERa, alfhangr t
to Wi, v'e
belong
Pe = P CzER2, ) F P(TEI, W2)

becomes

Pe PCreRa/w,) pewi)+RzERwz)Pw)

PDP 4s,) dz +P(u)| P(TAu) dz

BYes Rmle
sr uS1ng

Pe - P(/)p (r) d z t PCW/) pc) dr

is uow Casy to Sae fhat e erros is
minimizae
b the p r htionivig regons R snd R2 h e feamre

SP2ace ase cnosen So mat

vegi R 'Pw /2c) P(W/2)

>

reonR2 Paw2/2)> PCW l)

The e g o m s Ri, 22 covers al}
union f
SInie the
iatnihon
fram tha 4 a paka Lty
the space,
w a have fhatE
densiy fnchom
PCm/p(x) dz +
| PC/)p) de
- PC)

combn1ig
enahons O nd (

Pe PC) PCw)pcx) d+ PC/ )da-

dz
Pe PA)- (P()- PCW2/e)D pC)

Tnis snggestt That the pnhahy e r o s

f R, ls The egio spae m Slnl
mini' 2ed
Than Rz becomes the regm
PCw)> Plm/2)
1s t r e
Where the re
verse

a ro cos psblem
This he Cre G
elassifcanon trsk w u Melaseg
 W2 wMAn unknGun pakm, epRsi
bu hhe eamme vee.r s
A`sIgnd t chuw
PCwr/) >ÞCuf/2) S#i
So mimm2es tna elaskicaho ero

pretmbiik

Minimi3ng me classiicanon e r r o r preb<b

Cmipimigig The Averag Risk)
be the e o y fecthue spse Assi3ne4
Let R Assm
m d Ra o cas W2 A Pahure
e
tol a S s Wj
d a s s Wa Lis n
at belmgs Co
vecter z
miscass1fed m w,
t u i s v e e t s ts
R Then Commilpe
A penatu tern
s
End am erroT this r m do
WronA
W oaas
sll.
ASsociue b
A2 1S s o m a t Pe
seleehnig R, n d R2
Inslerd f will tre to
minimige a
we
mIimiged,
s ú e
mod
versl
0 dz+daPta)PCa
Jp(*o)
APCo) Ra
T=

sccordins to
each term is veghlid
ere
euosen seh tuat
sanihcanca A2 nd Aa
Are

d e c i s I MS afe penaugee u e
orreet

es Man

A eakd sn the ereet Aecs|tiy

 5

T h e aSSoeATue deesim eeArVes SF(-9 a ) o

are adncs Cie ellipsO1ds, paabolns,hyperbelas

pins imes)
elass1fr6
Io Sncb cases Hhe Bayesm
uat the
c l a s e r , m he
sense

nadraht
Fextue spRee
ls r med Vio
perfermed vio
xrhhon fhe
tos > 2 he
sulfnc
quad rabc decs bn
mparqnadrgEs.
are
swus
decis
decisiUn eurves are2
uadhe

Shonn belaw