d, 2004. spain), january 13r roc analysis icos y ... · 1 classifier evaluation in data mining: roc...

44
1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete (Spain), January 13rd, 2004. Dpto. de Sistemas Informáticos y Computación, Universidad Politécnica de Valencia, [email protected]

Upload: others

Post on 28-Jun-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

1

Cla

ssif

ier

Evalu

ati

on

in D

ata

Min

ing

:

RO

C A

naly

sis

José

Her

nánd

ez-O

rallo

Alb

acet

e(S

pain

), Ja

nuar

y 13

rd, 2

004.

Dpt

o. d

e Si

stem

as In

form

átic

os y

Com

puta

ción

,U

nive

rsid

ad P

olité

cnic

a de

Val

enci

a,

jora

llo@

dsic

.upv

.es

Page 2: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

2

Out

line

•In

trodu

ctio

n. T

he C

lass

ifica

tion

Task

and

its

Eva

luat

ion

•S

kew

-sen

sitiv

e E

valu

atio

n

•R

OC

Ana

lysi

s fo

r Cris

p C

lass

ifier

s

•R

OC

Ana

lysi

s fo

r Sof

t Cla

ssifi

ers

•Th

e A

UC

Met

ric: t

he A

rea

Und

er th

e R

OC

Cur

ve

•R

elat

ion

betw

een

AU

C a

nd E

rror

. Thr

esho

ld c

hoic

e.

•A

pplic

atio

ns

•E

xten

sion

to M

ore

than

Tw

o C

lass

es

•C

oncl

usio

ns

Page 3: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

3

Intro

duct

ion.

The

Cla

ssifi

catio

nTa

skan

dits

Eva

luat

ion

•C

lass

ifica

tion.

–O

ne o

f the

mos

t im

porta

nt ta

sks

in d

ata

min

ing.

Goa

l: to

obt

ain

a m

odel

, pat

tern

or f

unct

ion

that

tells

be

twee

n tw

o or

mor

e ex

clus

ive

clas

ses.

•C

lass

ifica

tion

Eva

luat

ion.

–Tr

aditi

onal

Mea

sure

for E

valu

atin

g C

lass

ifier

s:•

Err

or (a

lso

inve

rsel

y ac

cura

cy):

perc

enta

ge o

f bad

ly

clas

sifie

d in

stan

ces

(wrt.

the

test

set

or u

sing

cro

ss-

valid

atio

n / b

oots

trapp

ing)

.

Page 4: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

4

Intro

duct

ion.

The

Cla

ssifi

catio

nTa

skan

dits

Eva

luat

ion

•A

cla

ssifi

erpr

ovid

essu

ppor

tin

deci

sion

mak

ing

(bet

wee

ndi

ffere

ntac

tions

).

Are

we

still

mak

ing

deci

sion

sin

an

unsc

ient

ific

way

?

•Th

is q

uest

ion

is ra

ised

by:

–Sw

ets,

J.A

., D

awes

, R.M

., &

Mon

ahan

, J. (

2000

). “B

ette

r dec

isio

ns th

roug

h sc

ienc

e”Sc

ient

ific

Amer

ican ,

283

, 82-

87.

Page 5: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

5

Ske

w-s

ensi

tive

Eva

luat

ion

•D

istri

butio

n-se

nsiti

ve E

valu

atio

n:–

The

clas

ses

of th

e pr

oble

m d

o no

t hav

e th

e sa

me

dist

ribut

ion

(they

are

not

wel

l bal

ance

d, e

.g. 5

0% e

ach

of th

em)

•E

xam

ple:

–W

e ha

ve s

ever

al c

lass

ifier

s c 1

, c2,

c 3th

at p

redi

ct w

heth

er a

re

frige

ratio

n va

lve

has

to b

e op

ened

or c

lose

d in

a n

ucle

ar

plan

t.–

In o

rder

to e

valu

ate

the

clas

sifie

rs w

e us

e a

data

set o

btai

ned

durin

g la

st m

onth

, whe

re a

n op

erat

or h

as b

een

deci

ding

to

open

or t

o cl

ose

the

valv

e.•

100,

000

exam

ples

, fro

m w

hich

99,

500

are

of c

lass

“Clo

se”a

nd

500

are

of c

lass

“Ope

n”.

–Le

t’s s

ay th

at c

lass

ifier

c2

alw

ays

pred

ict “

Clo

se”(

trivi

al

clas

sifie

r).

•E

rror

of c

2: 0.

5%.

Isth

isa

good

clas

sifie

r?

Page 6: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

6

Ske

w-s

ensi

tive

Eva

luat

ion

•C

onfu

sion

/con

tinge

ncy

Mat

rix(e

.g. f

orth

ete

stse

t):Ac

tual

TNFN

CLO

SE (N

)FP

TPO

PEN

(P)

clos

e(n)

open

(p)

Dia

gona

l of

hits

Pred

icte

d

•Fr

omhe

re, s

ever

alm

etric

sha

vebe

ende

fined

:–

Pr(

P|p

) ≈Tr

ue P

ositi

ve R

ate:

TP

R =

TP

/ (T

P +

FN

). (“r

ecal

l”ot

”sen

sitiv

ity”o

t “po

sitiv

e ac

cura

cy”).

–P

r(N|p

) ≈Fa

lse

Neg

ativ

e R

ate:

FN

R =

FN

/ (T

P +

FN

). (“p

ositi

ve e

rror”)

–P

r(N|n

) ≈Tr

ue N

egat

ive

Rat

e: T

NR

= T

N /

(TN

+ F

P).

(”spe

cific

ity”o

t ”ne

gativ

e ac

cura

cy”).

–P

r(P

|n) ≈

Fals

e P

ositi

ve R

ate:

FP

R =

FP

/ (T

N +

FP

). (“

nega

tive

erro

r”)–

Pr(

p|P

) ≈P

ositi

ve P

redi

ctiv

e V

alue

: PP

V =

TP

/ (T

P +

FP

). (”p

reci

sion

”).

–P

r(n|N

) ≈N

egat

ive

Pre

dict

ive

Val

ue: N

PV

= T

N /

(TN

+ F

N).

–M

acro

-ave

rage

= A

VG

(TP

R, T

NR

). (T

hem

ean

can

be a

rithm

etic

, geo

met

ricor

othe

r)

–B

RE

AK

-EV

EN

=(P

reci

sion

+ R

ecal

l) / 2

= (P

PV

+ T

PR

) / 2

–F-

ME

AS

UR

E=

(Pre

cisi

on* R

ecal

l) / B

RE

AK

-EV

EN

= 2

*PP

V*T

PR

/ (P

PV

+ T

PR

)

Page 7: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

7

•E

xam

ple:

(te

stda

tase

twith

100,

000

inst

ance

s)

9900

020

0C

LOSE

500

300

OPE

Ncl

ose

open

c 1

Actu

al

Pred

.94

100

100

CLO

SE54

0040

0O

PEN

clos

eop

enc 3

Actu

al

9950

050

0C

LOSE

00

OPE

Ncl

ose

open

c 2

Actu

al

Ske

w-s

ensi

tive

Eva

luat

ion

ERR

OR

: 0,7

%

TPR

= 30

0 / 5

00 =

60%

FNR

=20

0 / 5

00 =

40%

TNR

= 99

000

/ 995

00 =

99,

5%FP

R=

500

/ 995

00 =

0.5

%PP

V= 3

00 /

800

= 37

.5%

NPV

= 99

000

/ 992

00 =

99.

8%

Mac

roav

g= (6

0 +

99.5

) / 2

=79

.75%

ERR

OR

: 0,5

%

TPR

= 0

/ 500

= 0

%FN

R=

500

/ 500

= 1

00%

TNR

= 99

500

/ 995

00 =

100

%FP

R=

0 / 9

9500

= 0

%PP

V= 0

/ 0

= IN

DEF

INID

ON

PV=

9950

0 / 1

0000

= 9

9.5%

Mac

roav

g= (0

+ 1

00 )

/ 2 =

50%

ERR

OR

: 5,5

%

TPR

= 40

0 / 5

00 =

80%

FNR

=10

0 / 5

00 =

20%

TNR

=94

100

/ 995

00 =

94.

6%FP

R=

5400

/ 99

500

= 5.

4%PP

V= 4

00 /

5800

= 6

.9%

NPV

= 94

100

/ 942

00 =

99.

9%

Mac

roav

g= (8

0 +

94.6

) / 2

=87

.3%

SpecificitySensitivity Recall Precision

Whi

chcl

assi

fieri

sbe

st?

Page 8: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

8

Ske

w-s

ensi

tive

Eva

luat

ion

•C

ost-s

ensi

tive

Eva

luat

ion:

–In

man

yci

rcum

stan

ces,

not

allt

heer

rors

prod

uced

by a

pr

edic

tive

mod

elha

veth

esa

me

cons

eque

nces

:•

Exa

mpl

e: k

eepi

nga

valv

ecl

osed

in a

nuc

lear

pla

ntw

hen

itsh

ould

be o

pen,

can

pro

voke

anex

plos

ion,

whi

leop

enin

ga

valv

ew

hen

itsh

ould

be c

lose

d, c

an p

rovo

kea

stop

.

–C

ostm

atrix

:

–Th

eim

porta

ntth

ing

isno

tto

obta

inth

e“c

lass

ifier

” with

few

er

erro

rs b

ut th

e on

e w

ith lo

wes

t cos

t.–

From

this

mat

rix, w

eca

n co

mpu

te th

eco

stof

a cl

assi

fier.

•C

lass

ifier

sar

e ev

alua

ted

with

thes

eco

sts.

•Th

ecl

assi

fierw

ithlo

wes

tcos

tis

chos

en.

Actu

al

020

00€

CLO

SE10

0€0

OPE

Ncl

ose

open

Pred

icte

d

Page 9: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

9

Ske

w-s

ensi

tive

Eva

luat

ion

•E

xam

ples

: Actu

alAc

tual

Actu

alC

onfu

sion

Mat

rices 0

2000

€C

LOSE

100€

0O

PEN

clos

eop

en

Actu

al

Pred

icte

d

9900

020

0C

LOSE

500

300

OPE

Ncl

ose

open

c 1

Pred

9410

010

0C

LOSE

5400

400

OPE

Ncl

ose

open

c 3

9950

050

0C

LOSE

00

OPE

Ncl

ose

open

c 2

Cos

tM

atrix

0€40

0,000

€C

LOSE

50,00

0€0€

OPE

Ncl

ose

open

c 1

0€20

0,000

€C

LOSE

540,0

00€

0€O

PEN

clos

eop

enc 3

0€1,0

00,00

0€C

LOSE

0€0€

OPE

Ncl

ose

open

c

Res

ultin

gM

atric

es

2

TOTA

L C

OST

: 450

,000

€TO

TAL

CO

ST: 1

,000

,000

€TO

TAL

CO

ST: 7

40,0

00€

Page 10: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

10

Ske

w-s

ensi

tive

Eva

luat

ion

•W

hata

ffect

sth

efin

al c

ost?

–Fo

rtw

ocl

asse

s. It

depe

nds

ona

cont

exto

rske

w:

•Th

eco

stof

fals

epo

sitiv

esan

dfa

lse

nega

tives

: FP

cost

and

FNco

st•

The

perc

enta

geof

exam

ples

ofth

ene

gativ

ecl

ass

wrt.

the

exam

ples

ofth

epo

sitiv

ecl

ass.

(Neg

/ Pos

).•

We

com

pute

: (fo

rthe

prev

ious

exam

ples

)

20120

0010

0=

=FN

cost

FPco

st19

950

099

500=

=Po

sN

eg1

199

9.95

20sl

ope=

×=

–Fo

rtw

ocl

asse

s, th

eva

lue

“slo

pe” i

s su

ffici

ent t

o de

term

ine

whi

ch c

lass

ifier

is b

est.

Cla

ssif.

1: F

NR

= 40

%, F

PR=

0.5%

Cos

t per

uni

t =

1 x

0.40

+ 9

.95

x 0.

005

= 0.

45

Cla

ssif.

2: F

NR

= 10

0%, F

PR=

0%C

ost p

er u

nit =

1

x 1

+ 9.

95x

0 =

1

Cla

ssif.

3: F

NR

= 20

%, F

PR=

5.4%

Cos

t per

uni

t =

1 x

0.20

+ 9

.95

x 0.

054

= 0.

74

Page 11: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

11

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•Th

e cl

assi

fier w

ith lo

wes

t err

or ra

te is

freq

uent

ly n

ot th

e be

st c

lass

ifier

.•

The

cont

exto

r ske

w(th

e cl

ass

dist

ribut

ion

and

the

cost

sof

eac

h er

ror)

det

erm

ine

the

good

ness

of

clas

sifie

rs.

•P

RO

BLE

M:

–In

man

y ci

rcum

stan

ces,

unt

il th

e ap

plic

atio

n tim

e, w

e do

not

kn

ow th

e cl

ass

dist

ribut

ion

and/

or it

is d

iffic

ult t

o es

timat

e th

e co

st m

atrix

. E.g

. a s

pam

filte

r.–

But

mod

els

are

lear

ned

befo

re.

•R

OC

(Rec

eive

r Ope

ratin

g C

hara

cter

istic

) Ana

lysi

s.–

Firs

t use

d in

the

WW

2 to

eva

luat

e ra

dars

, lat

er w

as u

sed

to s

tudy

the

resp

onse

of t

rans

isto

rs, a

nd fu

rther

dev

elop

ed fo

r med

ical

dia

gnos

is

appl

icat

ions

from

the

seve

ntie

s. It

beg

ins

to b

e kn

own

in d

ata

min

ing

in th

e la

te n

inet

ies.

Page 12: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

12

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•Th

eR

OC

Spa

ce–

Eac

hco

lum

nof

the

conf

usio

nm

atrix

isno

rmal

ised

: TP

R, F

NR

TN

R, F

PR

.Ac

tual

Actu

al

ROC

Spac

e

0,00

0

0,20

0

0,40

0

0,60

0

0,80

0

1,00

0

0,00

00,

200

0,40

00,

600

0,80

01,

000

Fals

e Po

sitiv

es

True Positives

8750

010

0C

LOSE

1200

040

0O

PEN

clos

eop

en

0.879

0.2C

LOSE

0.121

0.8O

PEN

clos

eop

en

TPR

= 40

0 / 5

00 =

80%

FNR

=10

0 / 5

00 =

20%

TNR

= 87

500

/ 995

00 =

87.

9%FP

R=

1200

0 / 9

9500

= 1

2.1%

Pred

Pred

Page 13: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

13

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•R

OC

spa

ce: g

ood

and

bad

clas

sifie

rs.

01

1 0FP

R

TPR •

Goo

dcl

assi

fier.

–H

igh

TPR

.–

Low

FPR

.

01

1 0FP

R

TPR

01

1 0FP

R

TPR •

Bad

clas

sifie

r.–

Low

TPR

.–

Hig

hFP

R.

•B

adcl

assi

fier

(rea

l pic

ture

).

Page 14: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

14

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•Th

eR

OC

Cur

ve R

OC

. “C

ontin

uity

”.R

OC

dia

gram

01

1 0FP

R

TPR

We

can

cons

truct

any

“in

term

edia

te” c

lass

ifier

just

ra

ndom

ly w

eigh

ting

both

cl

assi

fiers

(giv

ing

mor

e or

less

w

eigh

t to

one

or th

e ot

her).

This

cre

ates

a “c

ontin

uum

” of

clas

sifie

rs b

etw

een

any

two

clas

sifie

rs.

Giv

en tw

o cl

assi

fiers

:

Page 15: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

15

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•R

OC

Cur

ve. C

onst

ruct

ion.

RO

C d

iagr

am

01

1 0FP

R

TPR

We

cons

truct

the

conv

ex h

ull o

f th

eir p

oint

s (F

PR,T

PR) a

s w

ell a

s th

e tw

o tri

vial

cla

ssifi

ers

(0,0

) and

(1

,1).

The

clas

sifie

rs b

elow

the

RO

C

curv

e ar

e di

scar

ded.

The

best

cla

ssifi

er (f

rom

thos

e re

mai

ning

) will

be s

elec

ted

in

appl

icat

ion

time…

Giv

en s

ever

al c

lass

ifier

s:

We

can

disc

ard

thos

e w

hich

are

bel

ow b

ecau

se th

ere

is n

o co

mbi

natio

n of

cla

ss

dist

ribut

ion

/ cos

t mat

rix fo

r whi

ch th

ey c

ould

be

optim

al.

The

diag

onal

sh

ows

the

wor

st

situ

atio

n po

ssib

le.

Page 16: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

16

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•In

the

cont

exto

fapp

licat

ion,

we

choo

seth

eop

timal

clas

sifie

rfro

mth

ose

kept

. Exa

mpl

e1:

0%20%

40%

60%

80%

100%

0%20

%40

%60

%80

%10

0%

fals

e po

sitiv

e ra

te

true positive rate

FPco

stFN

cost=

1 2

Neg Po

s=

4

slop

e=

42=

2

Con

text

(ske

w):

Page 17: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

17

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•In

the

cont

exto

fapp

licat

ion,

we

choo

seth

eop

timal

clas

sifie

rfro

mth

ose

kept

. Exa

mpl

e2:

0%20%

40%

60%

80%

100%

0%20

%40

%60

%80

%10

0%

fals

e po

sitiv

e ra

te

true positive rate

FPco

stFN

cost=

1 8

Neg Po

s=

4

slop

e=

48=

.5

Con

text

(ske

w):

Page 18: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

18

RO

C A

naly

sis

ofC

risp

Cla

ssifi

ers

•W

hath

ave

we

lear

ned

from

this

?–

The

optim

ality

ofa

clas

sifie

rdep

ends

onth

ecl

ass

dist

ribut

ion

and

the

erro

r cos

ts.

–Fr

omth

isco

ntex

t/ s

kew

we

can

obta

inth

e“s

lope

”, w

hich

cha

ract

eris

es th

is c

onte

xt.

•If

we

know

this

cont

ext,

we

can

sele

ctth

ebe

st c

lass

ifier

, m

ultip

lyin

gth

eco

nfus

ion

mat

rixan

dth

eco

stm

atrix

.•

Ifw

edo

n’t k

now

this

con

text

in th

e le

arni

ng s

tage

, by

usin

g R

OC

ana

lysi

s w

e ca

n ch

oose

a s

ubse

t of c

lass

ifier

s, fr

om

whi

ch th

e op

timal

cla

ssifi

er w

ill b

e se

lect

ed w

hen

the

cont

ext

is k

now

n. Can

we

gofu

rther

than

this

?

Page 19: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

19

RO

C A

naly

sis

ofS

oftC

lass

ifier

s

•C

risp

and

Sof

tCla

ssifi

ers:

–A

“cris

p” c

lass

ifier

pre

dict

s a

clas

s be

twee

n a

set o

f pos

sibl

e cl

asse

s.–

A “s

oft”

clas

sifie

r (pr

obab

ilistic

) pre

dict

s a

clas

s, b

ut

acco

mpa

nies

eac

h pr

edic

tion

with

an

estim

atio

n of

the

relia

bilit

y (c

onfid

ence

) of e

ach

pred

ictio

n.•

Mos

tlea

rnin

gm

etho

dsca

n be

ada

pted

toge

nera

teth

isco

nfid

ence

valu

e.

•A

spe

cial

kind

ofso

ftcl

assi

fieri

sa

clas

spr

obab

ility

estim

ator

.–

Inst

ead

ofpr

edic

ting

“a”,

“b” o

r “c”

, it g

ives

a p

roba

bilit

y es

timat

ion

for “

a”, “

b” o

r “c”

, i.e

., “p

a”, “

p b” a

nd “p

c”. E

xam

ple:

•C

lass

ifier

1: p

a=0.

2, p

b=0.

5 an

dp c

=0.

3.•

Cla

ssifi

er2:

pa=

0.3,

pb=

0.4

and

p c=

0.3.

–B

oth

pred

ictb

, but

clas

sifie

r1 is

“sur

er”.

Page 20: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

20

RO

C A

naly

sis

ofS

oftC

lass

ifier

s

•“R

anke

rs”:

–W

hene

verw

eha

vea

prob

abili

tyes

timat

orfo

ra tw

o-cl

ass

prob

lem

:•

p a=

x, th

enp b

= 1 −

x.–

Itis

only

nece

ssar

yto

spec

ifyth

epr

obab

ility

ofon

ecl

ass.

–Le

t’s c

all o

ne c

lass

0 (n

eg) a

nd th

e ot

her c

lass

1 (p

os).

–A

rank

eris

a so

ftcl

assi

fiert

hatg

ives

a va

lue

betw

een

0 an

d1

ofth

epr

obab

ility

ofcl

ass

1. T

his

valu

eis

also

calle

d“s

core

” and

de

term

ines

whe

ther

the

pred

ictio

n is

clo

ser t

o cl

ass

0 or

cla

ss 1

.–

Exa

mpl

es:

•P

roba

bilit

yof

a cu

stom

erbu

ying

a pr

oduc

t.•

Pro

babi

lity

ofa

mes

sage

bein

gsp

am.

•...

Page 21: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

21

RO

C A

naly

sis

ofS

oftC

lass

ifier

s

•R

OC

Cur

ve o

fa S

oftC

lass

ifier

:–

A s

oftc

lass

ifier

can

be c

onve

rted

into

a cr

isp

clas

sifie

rusi

nga

thre

shol

d.•

Exa

mpl

e: “i

f sco

re >

0.7

then

cla

ss A

, oth

erw

ise

clas

s B

”.•

With

diffe

rent

thre

shol

ds, w

eha

vedi

ffere

ntcl

assi

fiers

, gi

ving

mor

e or

less

rele

vanc

eto

each

ofth

ecl

asse

s(w

ithou

tnee

dof

over

sam

plin

gan

dun

ders

ampl

ing)

.–

We

can

cons

ider

each

thre

shol

das

a d

iffer

ent

clas

sifie

rand

draw

them

in th

eR

OC

spa

ce. T

his

gene

rate

sa

curv

e…

We

have

a “c

urve

” for

just

one

sof

t cla

ssifi

er

•Th

iscu

rve

islik

ea

stai

r(th

eco

nvex

hull

isno

tdon

e ge

nera

lly).

Page 22: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

22

RO

C A

naly

sis

ofS

oftC

lass

ifier

sA

ctua

l Cla

ss

•R

OC

Cur

ve o

fa S

oftC

lass

ifier

:–

Exa

mpl

e:n n n n n n n n n n n n n n n n n n n n

Pred

icte

dC

lass p p p p p p p p p p p p p p p p p p p p

p n n n n n

n n n

n n n

n n n

n n n

n n

p p n n n n

n n n

n n n

n n n

n n n

n n

...

© T

omFa

wce

tt

Page 23: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

23

RO

C A

naly

sis

ofS

oftC

lass

ifier

s

•R

OC

Cur

ve o

fa S

oftC

lass

ifier

:

© T

omFa

wce

tt

Page 24: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

24

RO

C A

naly

sis

ofS

oftC

lass

ifier

s

•R

OC

Ana

lysi

sof

seve

ral“

soft”

cla

ssifi

ers:

In th

iszo

neth

ebe

st

clas

sifie

ris

“inst

s”

In th

iszo

neth

ebe

st

clas

sifie

ris“

inst

s2”

•W

em

ustp

rese

rve

the

clas

sifie

rsth

atha

veat

leas

tone

“bes

t zon

e” a

nd

then

beh

ave

in th

e sa

me

way

as

we

did

for c

risp

clas

sifie

rs.

© R

ober

tHol

te

Page 25: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

25

The

“AU

C” M

etric

: the

Are

a U

nder

the

RO

C C

urve

RO

C c

urve

0,00

0

0,20

0

0,40

0

0,60

0

0,80

0

1,00

0

0,00

00,

200

0,40

00,

600

0,80

01,

000

Fals

e Po

sitiv

es

True Positives

•W

hati

fwe

wan

tto

sele

ctju

ston

ecl

assi

fier?

–Th

ecl

assi

fierw

ithgr

eate

stAr

eaU

nder

the

RO

C C

urve

(A

UC

) is

chos

en.

AU

C

–Fo

rcris

pcl

assi

fiers

itis

equi

vale

ntto

the

mac

roav

erag

e.

Alte

rnat

ive

toer

ror f

orev

alua

ting

clas

sifie

rs•

A d

ata

min

ing

/ lea

rnin

gm

etho

dw

illbe

bet

teri

fit

gene

rate

scl

assi

fiers

with

high

AUC

.

Page 26: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

26

The

“AU

C” M

etric

: the

Are

a U

nder

the

RO

C C

urve

•W

hati

fwe

wan

tto

sele

ctju

ston

eso

ftcl

assi

fier?

–Th

ecl

assi

fierw

ithgr

eate

stAr

eaU

nder

the

RO

C C

urve

(A

UC

) is

chos

en.

In th

isca

se w

ese

lect

B.

© T

omFa

wce

tt

But

fort

he“s

oft”

case

we

have

sur

pris

es…

Page 27: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

27

The

“AU

C” M

etric

: the

Are

a U

nder

the

RO

C C

urve

•Th

e A

UC

and

the

Wilc

oxon

-Man

n-W

hitn

ey (W

MW

) (W

ilcox

on19

45) (

Man

n &

Whi

tney

194

7) s

tatis

tic a

re e

quiv

alen

t .–

The

WM

W te

st is

use

ful t

o de

term

ine

whe

ther

one

of t

wo

rand

om

varia

bles

is s

toch

astic

ally

gre

ater

than

the

othe

r.P[

X>Y]

–If

we

choo

se X

as th

e ex

ampl

es o

f one

cla

ss a

nd Y

as th

e ex

ampl

es o

f th

e ot

her c

lass

, and

the

valu

es X

and

Yas

the

estim

ated

sco

reby

the

clas

sifie

r, w

e ha

ve th

at th

e A

UC

is e

quiv

alen

t to

the

WM

W s

tatis

tic.

Sho

uld

I be

happ

yof

this

?Th

eA

UC

real

lyes

timat

esth

epr

obab

ility

that

, ifw

ech

oose

anex

ampl

eof

clas

s1

and

anex

ampl

eof

clas

s0,

the

clas

sifie

rwill

give

mor

e sc

ore

toth

efir

ston

eth

anto

the

seco

ndon

e.

(Car

e! T

his

does

notm

ean

that

itcl

assi

fies

wel

lbot

hex

ampl

es).

But:

•Th

ere

does

exis

ta th

resh

old

from

whi

chit

can

clas

sify

both

corre

ctly

.•

Itca

nnot

clas

sify

both

wro

ngly

, for

any

thre

shol

dw

hats

oeve

r.

Page 28: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

28

The

“AU

C” M

etric

: the

Are

a U

nder

the

RO

C C

urve

•Th

e A

UC

met

ric fo

r sof

t cla

ssife

rsor

rank

ers.

–E

valu

ates

how

wel

l a c

lass

ifier

per

form

s a

rank

ing

of it

s pr

edic

tions

.or

, in

othe

r wor

ds,

–E

valu

ates

how

wel

l a c

lass

ifier

is a

ble

to s

ort i

ts p

redi

ctio

ns

acco

rdin

g to

the

conf

iden

ce it

ass

igns

to th

em.

–Th

e “r

anki

ngs”

of p

redi

ctio

ns a

re fu

ndam

enta

l in

man

y ap

plic

atio

ns:

•Fr

aud

dete

ctio

n.•

Mai

ling

cam

paig

n de

sign

.•

Spa

m fi

lterin

g.•

Failu

re d

etec

tion,

med

ical

dia

gnos

is.

•…

–A

nd m

any

othe

r dat

a m

inin

g m

etho

ds:

•C

ombi

natio

n of

cla

ssifi

ers.

•C

olla

bora

tive

Met

hods

. Rec

omm

ende

r sys

tem

s…

Page 29: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

29

Rel

atio

nbe

twee

nA

UC

and

erro

r. Th

resh

old

choi

ce

•W

eha

vese

enth

atA

UC

isa

bette

reva

luat

ion

mea

sure

than

erro

r rat

e. B

ut, w

hich

rela

tion

isth

ere

betw

een

both

?–

Logi

cally

, an

AU

C c

lose

rto

1 w

illgi

vean

erro

r rat

ecl

ose

to0.

01

1 0FP

R

TPR

•W

hate

vers

kew

we

have

, a lo

wer

ror i

sex

pect

ed.

9950

050

0C

LOSE

00

OPE

NC

lose

open

c 2

01

1 0FP

R

TPR

–B

utan

erro

r rat

ecl

oser

to0

does

note

nsur

ean

AU

C c

lose

to1.

•Le

t’s re

call

the

nucl

ear p

lant

exa

mpl

e:

AUC

= 0

.5(M

inim

um p

ossi

ble

valu

e)M

acro

avg=

AU

C =

(0

+ 1

00 )

/ 2 =

50%

ERR

OR

: 0.

005

Page 30: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

30

Rel

atio

nbe

twee

nA

UC

and

erro

r. Th

resh

old

choi

ce

•M

any

lear

ning

met

hods

are

soft

by d

efin

ition

and

are

conv

erte

din

tocr

isp

forp

erfo

rmin

gth

ecl

assi

ficat

ion.

–Fo

rexa

mpl

e, a

Naï

veB

ayes

cla

ssifi

er, f

ora

two-

clas

spr

oble

m(a

and

b), e

stim

ates

two

prob

abili

ties:

P(a|

x) a

ndP(

b|x)

.–

The

clas

sific

atio

nru

leis

the

follo

win

gon

e:

–Th

isru

le“w

aste

s” a

Bay

esia

n cl

assi

fier.

But

, whi

ch o

ther

rule

co

uld

be u

sed?

(Lac

hich

e &

Fla

ch 2

003)

•Fi

rst,

we

conv

ertt

hepr

obab

ilitie

sin

tosc

ores

, as

follo

ws.

s(x)

= P

(a|x

) / P

(b|x

)

IfP(

a|x)

> P

(b|x

) the

ncl

ass

aO

ther

wis

ecl

ase

b

Page 31: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

31

Rel

atio

nbe

twee

nA

UC

and

erro

r. Th

resh

old

choi

ce

•N

ow, s

impl

y, le

t’s d

o R

OC

ana

lysi

s.–

We

draw

the

RO

C c

urve

usi

ngth

esc

ore.

–W

eco

mpu

te th

esl

ope

from

the

a pr

iori

clas

sdi

strib

utio

n(o

rthe

dist

ribut

ion

inap

plic

atio

ntim

e, a

ndco

sts

ifw

eha

veth

em).

–W

ech

oose

the

thre

shol

dbe

twee

nbo

thcl

asse

sfo

rs(

x).

© N

icol

asLa

chic

he&

Pet

erFl

ach

Cho

ice

usin

g ru

le

P(a|

x) >

P(b

|x)

Cho

ice

usin

g R

OC

ana

lysi

s an

d or

igin

al

dist

ribut

ion

Slop

egi

ven

by

the

orig

inal

di

strib

utio

n(tr

ain

orte

st).

Page 32: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

32

Rel

atio

nbe

twee

nA

UC

and

erro

r. Th

resh

old

choi

ce

•E

xam

ple

ofre

sults

(acc

urac

y) fo

r25

data

sets

:(L

achi

che

& F

lach

2003

)

–R

esul

tsar

e im

prov

edby

ju

stch

angi

ngth

ede

cisi

onru

leof

the

Naï

veB

ayes

C

lass

ifier

.

© N

icol

asLa

chic

he&

Pet

erFl

ach

Page 33: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

33

Rel

atio

nbe

twee

nA

UC

and

erro

r. Th

resh

old

choi

ce

•In

the

othe

r way

roun

d, it

wou

ld b

e in

tere

stin

g to

co

nver

t dat

a m

inin

g m

etho

ds th

at o

btai

n cr

isp

clas

sifie

rs in

to s

oft c

lass

ifier

s.–

This

wou

ld a

llow

its

use

as ra

nker

s, fo

r com

bina

tion,

...

–Th

is w

ould

mak

e it

poss

ible

to c

hoos

e a

bette

r thr

esho

ld a

nd

impr

ove

resu

lts.

•M

any

clas

sica

l met

hods

hav

e be

en re

desi

gned

and

re-

thou

ght.

For i

nsta

nce,

for d

ecis

ion

trees

:–

Dec

isio

n tre

es a

re le

arne

d us

ing

AU

C a

s sp

littin

g cr

iterio

n (F

erri

et a

l. 20

02).

–Th

e le

af p

roba

bilit

ies

are

smoo

thed

(usi

ng L

apla

ce c

orre

ctio

n or

oth

er m

ore

soph

istic

ated

sm

ooth

ing

met

hods

) and

pru

ning

is

show

n to

be

coun

terp

rodu

ctiv

e to

obt

ain

good

AU

C m

easu

res

(Pro

vost

& D

omin

gos

2003

) (Li

ng &

Yan

2003

) (Fe

rriet

al.

2003

a).

Page 34: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

34

App

licat

ions

•An

exam

ple:

“mai

l cam

paig

n de

sign

”:–

A co

mpa

nyw

ants

tom

ake

a m

aile

dof

fert

oin

crea

seth

esa

les

ofon

eof

itspr

oduc

ts. I

n ca

se o

fpos

itive

repl

ycu

stom

ers

buy

prod

ucts

with

a m

ean

valu

eof

100€

. If 5

5% a

re p

rodu

ctio

n co

sts

(fix

and

varia

ble)

, we

have

that

for e

ach

posi

tive

repl

y th

ere

is a

m

ean

gain

of 4

5€.

–Ea

chpo

st s

enth

as a

cos

tof1

€ (m

ail,

leaf

let)

and

the

who

le o

f th

e ca

mpa

ign

(inde

p. o

f the

num

ber o

f pos

tings

) wou

ld h

ave

a ba

se c

ost o

f 20,

000€

.–

With

1,00

0,00

0 cu

stom

ers,

forw

hich

, with

a fir

stsa

mpl

eof

1,00

0 cu

stom

ers,

we

have

estim

ated

that

1% re

plie

s(b

uys)

...

How

shou

ldw

eac

t?

Page 35: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

35

App

licat

ions

•“m

ail c

ampa

ign

desi

gn”:

–Le

t’s c

alcu

late

the

cost

s / b

enef

its.

•C

OST

S (a

ssum

ing

that

we

send

the

cam

paig

nto

allc

usto

mer

s):

–20

,000

€ de

sign

of t

he c

ampa

ign.

–1,

000,

000

x 1€

= 1

,000

,000

€.–

TOTA

L: 1

,020

,000

ۥ

BEN

EFIT

S–

1% o

frep

lies

from

1,00

0,00

0 ar

e 10

,000

repl

ies,

45€

eac

h on

e.–

TOTA

L: 4

50,0

00€

–M

ore

cost

sth

anbe

nefit

sC

ance

lled

cam

paig

n.

Hav

ew

edo

ne w

ell?

Page 36: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

36

•“m

ail c

ampa

ign

desi

gn”:

–Le

tus

train

a so

ftcl

assi

fierw

ithth

esa

mpl

ean

dle

tus

draw

itsR

OC

cur

ve (w

itha

prev

ious

lyre

mov

ed p

artf

orte

stse

t).

App

licat

ions

Cos

t Mat

rixBu

ys

0

0,1

0,2

0,3

0,4

0,5

0,6

0,7

0,8

0,91

00,

10,

20,

30,

40,

50,

60,

70,

80,

91

-44€

1€SÍ

-45€

0€N

Osí

no

Sent

But t

he c

onfu

sion

mat

rix h

as

an im

poss

ible

cell:

“Not

sen

t an

d bo

ught

•Ad

ditio

nally

, the

clas

sifie

ris

notv

ery

good

(clo

sert

oth

edi

agon

al)

tokn

ow w

hich

cust

omer

sw

illbu

yan

dw

hich

cust

omer

sw

on’t.

•Ev

enth

ough

, we

can

do m

any

thin

gs...

Page 37: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

37

App

licat

ions

•“m

ail c

ampa

ign

desi

gn”:

–W

eca

n us

e th

ecl

assi

fiert

ode

term

ine

who

tose

ndth

eof

fer.

Cos

tofC

ampa

ign

20,0

00 --

> 20

,000

100,

000

x 1

--> 1

00,0

00

Tota

l:12

0,00

0

Ben

ef. C

ampa

ign

3,00

0 x

45 --

> 13

5,00

0

Net

Ben

ef.:

1

5,00

0

Cos

tofC

ampa

ign

20,0

00 --

> 20

,000

200,

000

x 1

--> 1

00,0

00

Tota

l:22

0,00

0

Ben

ef. C

ampa

ign

5,00

0 x

45 --

> 22

5,00

0

Net

Ben

ef.:

5

,000

Page 38: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

38

App

licat

ions

•“m

ail c

ampa

ign

desi

gn”:

–G

raph

show

ing

the

bene

fitfo

rthr

eedi

ffere

ntca

mpa

igns

...

Page 39: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

39

Ext

ensi

onto

Mor

e th

anTw

oC

lass

es

•Th

ete

chni

ques

pres

ente

dhe

reha

vebe

enillu

stra

ted

for

two

clas

ses:

–Th

eev

alua

tion

ofm

ultic

lass

clas

sifie

rsba

sed

onco

sts

can

be

perfo

rmed

equa

lly.

•Ex

ampl

e:C

OST

actu

al

low

m

ediu

mhi

gh

low

0€

5€

2€

m

ediu

m20

0€

-200

0€

10€

pr

edic

ted

high

10

€ 1€

-1

5€

To

tal c

ost:

-297

87€

ERR

OR

actu

al

low

m

ediu

mhi

gh

low

20

0

13

med

ium

5 15

4

pr

edic

ted

high

4

7 60

Page 40: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

40

Ext

ensi

onto

Mor

e th

anTw

oC

lass

es

•R

OC

ana

lysi

s, th

ough

, is

not e

asily

ext

ensi

ble:

–G

iven

ncl

asse

s, th

ere

is a

nx

(n−1

) dim

ensi

onal

spa

ce.

•C

alcu

latin

g th

e co

nvex

hul

l im

prac

tical

.•

The

“con

text

” now

is n

ot d

eter

min

ed b

y a

valu

e (s

lope

), bu

t nx

(n−1

) −1

val

ues.

Ther

e ha

ve b

een

appr

oxim

atio

ns (f

or th

ree

clas

ses,

Mos

sman

199

9)or

ef

forts

to ta

ckle

the

gene

ral p

robl

em (S

riniv

asan

1999

) (Fe

rriet

al.

2003

b).

∑∑

=<

=−

=c i

c

ij

jH

Tj

iAU

Cc

cAU

C1

,1)

,()1

(1

•Th

e AU

C m

easu

re h

as b

een

exte

nded

.–

All-p

air e

xten

sion

(Han

d &

Till

2001

).

–“o

ne v

s. re

st” e

xten

sion

(Faw

cett)

–O

ther

ext

ensi

ons

(Yan

et a

l. 20

03) (

AUC

*, Ti

ng 2

002)

.

Page 41: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

41

Con

clus

ions

•R

OC

Ana

lysi

s:

–H

ighl

ight

s th

e fa

ct th

at th

e ev

alua

tion

of c

lass

ifier

s go

es m

uch

beyo

nd th

e es

timat

ion

of th

e er

ror r

ate.

–M

akes

it p

ossi

ble

to w

ork

with

cost

s an

d di

strib

utio

ns, o

r with

out

this

info

rmat

ion,

impr

ovin

g th

e ge

nera

tion,

sel

ectio

n an

d ap

plic

atio

n of

cla

ssifi

ers.

–P

rovi

des

a se

t of v

arie

d m

etric

s an

d te

chni

ques

to e

valu

ate

clas

sifie

rs d

epen

ding

on

the

task

: min

imis

eer

ror,

min

imis

eco

st,

impr

ove

a ra

nkin

g, e

tc.

•It

is a

hot

rese

arch

are

a of

wid

e ap

plic

abili

ty in

dat

a m

inin

g an

d m

achi

ne le

arni

ng.

Page 42: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

42

Som

eR

efer

ence

s

•Br

adle

y, A

.P. (

1997

) “Th

e us

e of

the

area

und

er th

e R

OC

cur

ve in

the

eval

uatio

n of

mac

hine

lear

ning

al

gorit

hms”

Patte

rn R

ecog

nitio

n, 3

0(7)

, 114

5-11

59.

•Eg

an, J

.P. (

1975

). Si

gnal

Det

ectio

n Th

eory

and

RO

C An

alys

is. S

erie

s in

Cog

nitio

n an

d Pe

rcep

tion.

Ac

adem

ic P

ress

, New

Yor

k.•

Faw

cett,

T.(2

001)

. “U

sing

rule

set

s to

max

imiz

e R

OC

per

form

ance

”.In

Proc

eedi

ngs

of th

e IE

EE In

tern

atio

nal

Con

fere

nce

on D

ata

Min

ing

(ICD

M-2

001)

, pp.

131-

138.

•Fa

wce

tt, T

., &

Prov

ost,

F. (1

997)

. “Ad

aptiv

e fra

ud d

etec

tion”

. Dat

a M

inin

g an

d Kn

owle

dge

Disc

over

y,

1(3)

,291

-316

.•

Faw

cett,

T. (2

003)

. “R

OC

gra

phs:

Not

es a

nd p

ract

ical

con

side

ratio

ns fo

r dat

a m

inin

g re

sear

cher

s”Te

ch

repo

rt H

PL-2

003-

4. H

P La

bora

torie

s, P

aloA

lto, C

A, U

SA. A

vaila

ble:

ht

tp://

ww

w.p

url.o

rg/n

et/tf

awce

tt/pa

pers

/HPL

-200

3-4.

pdf.

•Fe

rri, C

., Fl

ach,

P.;

Her

nánd

ez-O

rallo

, J. (

2002

). “L

earn

ing

Dec

isio

n Tr

ees

usin

g th

e Ar

ea U

nder

the

RO

C

Cur

ve”,

inC

. Sam

mut

; A. H

offm

an (e

ds.)

“The

200

2 In

tern

atio

nal C

onfe

renc

e on

Mac

hine

Lea

rnin

g”(IC

ML2

002)

, IO

S Pr

ess,

Mor

gan

Kauf

man

n Pu

blis

hers

, pp.

139

-146

.•

Ferri

, C.;

Flac

h, P

.A.;

Her

nánd

ez-O

rallo

, J. (

2003

a) "I

mpr

ovin

g th

e AU

C o

f Pro

babi

listic

Est

imat

ion

Tree

s".

Euro

pean

Con

fere

nce

on M

achi

ne L

earn

ing,

EC

ML

2003

: 121

-132

•Fe

rri, C

.; H

erná

ndez

-Ora

llo, J

.; Sa

lido,

M.A

. (20

03b)

"Vol

ume

unde

r the

RO

C S

urfa

ce fo

r Mul

ti-cl

ass

Prob

lem

s". E

urop

ean

Con

fere

nce

on M

achi

ne L

earn

ing,

EC

ML

2003

: 108

-120

•Fl

ach,

P.;

Bloc

keel

, H.;

Ferri

, C.;

Her

nánd

ez-O

rallo

, J.;

Stru

yf, J

. (20

03) “

Dec

isio

n Su

ppor

t for

Dat

a M

inin

g:

Intro

duct

ion

to R

OC

ana

lysi

s an

d its

app

licat

ions

”in

Data

Min

ing

and

Decis

ion

Supp

ort:

Inte

grat

ion

and

Colla

bora

tion ,

Klu

wer

Aca

dem

ic P

ublis

hers

, Bos

ton,

200

3.•

Flac

h, P

.A. "

The

Geo

met

ryof

RO

C S

pace

: Und

erst

andi

ngM

achi

neLe

arni

ngM

etric

sth

roug

hR

OC

Is

omet

rics"

. (20

03) I

nter

natio

nal C

onfe

renc

e on

Mac

hine

Lea

rnin

g, IC

ML

2003

: 194

-201

•Fü

rnkr

anz,

J.;

Flac

h, P

.A.:

AnAn

alys

isof

Rul

eEv

alua

tion

Met

rics.

(200

3) In

tern

atio

nal C

onfe

renc

e on

M

achi

ne L

earn

ing,

ICM

L 20

03: 2

02-2

09•

Han

d, D

.J.,

& Ti

ll, R

.J. (

2001

). “A

Sim

ple

Gen

eral

isat

ion

of th

e Ar

ea U

nder

the

RO

C C

urve

for M

ultip

le C

lass

C

lass

ifica

tion

Prob

lem

s”, M

achi

ne L

earn

ing,

45,

pp.

171

-186

.•

Han

ley,

J.A

. , &

McN

eil,

B.J.

(198

2). “

The

mea

ning

and

use

of t

he a

rea

unde

r a re

ceiv

er o

pera

ting

char

acte

ristic

(RO

C) c

urve

”. Ra

diol

ogy,

143,

29-3

6.

Page 43: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

43

Som

eR

efer

ence

s

•La

chic

he, N

. & F

lach

, P.A

. (20

03).

“Impr

ovin

g Ac

cura

cy a

nd C

ost o

f Tw

o-cl

ass

and

Mul

ti-cl

ass

Prob

abilis

tic

Cla

ssifi

ers

Usi

ng R

OC

Cur

ves”

. Int

erna

tiona

l Con

fere

nce

on M

achi

ne L

earn

ing,

ICM

L 20

03: 4

16-4

23•

Lane

, T. (

2000

). “E

xten

sion

s of

RO

C a

naly

sis

to m

ulti-

clas

s do

mai

ns”.

In D

iette

rich,

T.,

Mar

gine

antu

, D.,

Prov

ost,

F., &

Tur

ney,

P. (

Eds.

), IC

ML-

2000

Wor

ksho

p on

Cos

t-Sen

sitiv

e Le

arni

ng•

Ling

, C.X

.; Ya

n, R

.J. (

2003

) “D

ecis

ion

Tree

with

Bet

ter R

anki

ng”T

he 2

003

Inte

rnat

iona

l Con

fere

nce

on

Mac

hine

Lea

rnin

g (IC

ML2

003)

, IO

S Pr

ess,

Mor

gan

Kauf

man

n Pu

blis

hers

, to

appe

ar.

•M

ann,

H. B

. & W

hitn

ey, D

. R. (

1947

). "O

n a

test

whe

ther

one

of t

wo

rand

om v

aria

bles

is s

toch

astic

ally

larg

er

than

the

othe

r". A

nn. M

ath.

Sta

tist.,

18,

pp.

50-

60.

•M

ossm

an,D

.(199

9). “

Thre

e-w

ay R

OC

s”. M

edica

l Dec

ision

Mak

ing,

19,7

8-89

.•

Prov

ost,

F., &

Dom

ingo

s, P

. (20

03).

“Tre

e In

duct

ion

for P

roba

bilit

y-ba

sed

Ran

king

”, M

achi

ne L

earn

ing

52:3

(in

pre

ss),

2003

.•

Srin

ivas

an, A

. (19

99) “

Not

e on

the

Loca

tion

of O

ptim

al C

lass

ifier

s in

N-d

imen

sion

al R

OC

Spa

ce”T

echn

ical

R

epor

t PR

G-T

R-2

-99,

Oxf

ord

Uni

vers

ity C

ompu

ting

Labo

rato

ry, O

xfor

d.•

Swet

s, J

.A. (

1988

). “M

easu

ring

the

accu

racy

of d

iagn

ostic

sys

tem

s”. S

cienc

e, 2

40,1

285-

1293

.•

Swet

s, J

.A.,

Daw

es, R

.M.,

& M

onah

an, J

. (20

00).

“Bet

ter d

ecis

ions

thro

ugh

scie

nce”

. Sci

entif

ic A

mer

ican

, 28

3, 8

2-87

. http

://w

ww

.psy

chol

ogic

alsc

ienc

e.or

g/pd

f/psp

i/sci

am.p

df.

•Ti

ng, K

aiM

ing

(200

2). “

Issu

es in

Cla

ssifi

er E

valu

atio

n us

ing

Opt

imal

Cos

t Cur

ves”

The

Pro

ceed

ings

of t

he

Inte

rnat

iona

l Con

fere

nce

on M

achi

ne L

earn

ing"

Inte

rnat

iona

l Con

fere

nce

on M

achi

ne L

earn

ing,

ICM

L 20

02,

pp. 6

42-6

49.

•Tu

rney

, P. (

2000

) “Ty

pes

of C

osti

n In

duct

ive

Con

cept

Lea

rnin

g”Pr

ocee

ding

s W

orks

hop

on C

ost-S

ensi

tive

Lear

ning

at t

he S

even

teen

th In

tern

atio

nal C

onfe

renc

e on

Mac

hine

Lea

rnin

g (W

CSL

at I

CM

L-20

00),

15-2

1.•

Wei

ss, G

. and

Pro

vost

, F. “

The

Effe

ct o

f Cla

ss D

istri

butio

n on

Cla

ssifi

er L

earn

ing:

An

Empi

rical

Stu

dy”

Tech

nica

l Rep

ort M

L-TR

-44,

Dep

artm

ent o

f Com

pute

r Sci

ence

, Rut

gers

Uni

vers

ity, 2

001.

Wilc

oxon

, F. (

1945

). "In

divi

dual

com

paris

ons

by ra

nkin

g m

etho

ds".

Biom

etric

s, 1

, pp.

80-

83.

•Ya

n, L

., D

odie

r, R

., M

ozer

, M. C

., &

Wol

niew

icz,

R. (

2003

). "O

ptim

izin

g cl

assi

fier p

erfo

rman

ce v

ia th

e W

ilcox

on-M

ann-

Whi

tney

sta

tistic

. In

The

Pro

ceed

ings

of t

he In

tern

atio

nal C

onfe

renc

e on

Mac

hine

Lea

rnin

g"

Inte

rnat

iona

l Con

fere

nce

on M

achi

ne L

earn

ing,

ICM

L (p

p. 8

48-8

55).

http

://w

ww

.cs.

colo

rado

.edu

/~m

ozer

/pap

ers/

•Zw

eig,

M.H

.; C

ampb

ell,

G. (

1993

) “R

ecei

ver-o

pera

ting

char

acte

ristic

(RO

C) p

lots

: a fu

ndam

enta

l eva

luat

ion

tool

in c

linic

alm

edic

ine”

, Clin

. Che

m, 1

993;

39:

561

-77.

Page 44: d, 2004. Spain), January 13r ROC Analysis icos y ... · 1 Classifier Evaluation in Data Mining: ROC Analysis José Hernández-Orallo Albacete Spain), January 13r d, 2004. Dpto. d

44

Som

eW

ebsi

tes

toK

now

Mor

e

•To

mFa

wce

tt’s

page

on

RO

C A

naly

sis:

http

://w

ww

.hpl

.hp.

com

/per

sona

l/Tom

_Faw

cett/

RO

CC

H/

•R

OC

Ana

lysi

sSo

ftwar

eht

tp://

epiw

eb.m

asse

y.ac

.nz/

RO

C_a

naly

sis_

softw

are.

htm

http

://cs

.bris

.ac.

uk/~

farra

nd/ro

con/

•R

OC

Ana

lysi

s Ex

tens

ive

Bibl

iogr

aphy

:ht

tp://

splw

eb.b

wh.

harv

ard.

edu:

8000

/pag

es/p

pl/z

ou/ro

c.ht

ml

•1s

t Wor

ksho

p on

“RO

C A

naly

sis

in A

I”, V

alen

cia,

22/

23 A

ugus

t 20

04 (w

ithin

EC

AI’2

004)

http

://w

ww

.dsi

c.up

v.es

/~fli

p/R

OC

AI20

04/