29
8/12/2019 Novel Approaches Discrimination http://slidepdf.com/reader/full/novel-approaches-discrimination 1/29 ZOR - Methods and Models of Operations Research (1992) 36:517-545 Novel pproaches to the Discrimination Problem P. Marcotte and G. Savard 1 Abstract We consider the problem of determining a hyperplane that separates, as well as possible, two finite sets of points in R . We analyze two criteria for judging the quality of a candidate hyperplane (i) the maximal distance of a misclassitied point to the hyperplane (ii) the number of misclassified points. In each case, we investigate the computational complexity of the corresponding mathematical programs, give equivalent formulations, suggest solution algorithms and present preliminary numerical results. Key words Discriminant analysis. Quadratic programming. Complexity. Integer programming. Bilevel programming. Introduction Consider random samples X {xi}i~1 and Y --- J {Y }j~s from two distinct popu- lations 3f and qr in R (x i e 5f, yJ ~ q/) with respective distribution functions Fx and Ft. Given a randomly generated vector z originating from either population 5( or qr statistical linear discriminant analysis uses a hyperplane as a tool to decide whether z e &r or z e qr namely z e X if z lies above the hyperplane and z e qr if z lies below the hyperplane. Statistical discriminant analysis is an established field of statistical theory. In its simplest, linear form, it consists in determining the best hyperplane, i.e. the hyperplane that maximizes the prob- ability of correctly assigning z to population 5f or qr When {xi}i~i and J are random samples drawn from multivariate normal populations with respective means/~x and #r and common variance-covariance matrix 2~, it can be shown that the optimal hyperplane is given by the formula (see Fisher 14]): ptz a=O Research supported by NSERC grants 5789 and 46405, the Academic Research Program of the Department of National Defense (Canada) and FCAR grant 91NC0510. (Quebec). 1 Professors Dr. Patrice Marcotte and Gilles Savard, Drpartement de Math+matiques, Coll~ge militaire royal de Saint-Jean, Richelain Qurbec J0J 1R0, Canada. 0340- 9422/92/6/517- 545 $2.50 9 1992 Physica-Verlag, Heidelberg

Novel Approaches Discrimination

Embed Size (px)

Citation preview

Page 1: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 1/29

Z O R - M e t h o d s a n d M o d e l s o f O p e r a t io n s R e s e a rc h ( 19 9 2) 3 6 :5 1 7 - 5 4 5

N o v e l p p ro a ch es t o t h e D i s cr im i n a t io n P ro b l em

P . M a r c o t t e a n d G . S a v a r d 1

Abstract W e con sider the pro blem of determ ining a hype rplane that separates , as we l l as possible,

two f in i t e se t s o f po in t s i n R . We ana lyze two c r i t e r i a fo r j udg ing the qua l i t y o f a cand ida t e

hype rplane ( i ) the m axim al d is tance of a misclassit ied po int to th e hyperp lane ( i i ) the nu mb er o f

misc lassi fi ed po in t s . I n each case, we inves t iga t e t he com puta t iona l com plex i ty o f t he co r r espond ing

ma them at i ca l p rograms, g ive equ iva l en t f o rmula t ions , suggest so lu t ion a lgor i thms and p resen t

prel im inary nu me r ical results.

Key words Discr iminan t ana lys i s . Quadra t i c p rogramming . Complex i ty . I n t eger p rogramming .

Bi level programming.

Introduct ion

Co n s i d e r r a n d o m s am p l e s X {xi} i~1 an d Y --- J{Y } j ~s f ro m t w o d i s t i n c t p o p u -

l a t i o n s 3 f an d qr i n R (x i e 5 f, y J ~ q/ ) w i t h r e s p ec t iv e d i s t r i b u t i o n fu n c t i o n s F x

a n d F t . G i v e n a r a n d o m l y g e n e ra t e d v e c t o r z o r ig i n a t in g fr o m ei ther p o p u l a t i o n

5 ( o r qr s t a t i s ti c a l l i n ea r d i s c r i m i n an t an a l y s i s u s e s a h y p e r p l an e a s a t o o l t o

d ec i d e w h e t h e r z e &r o r z e qr n am e l y z e X i f z l ie s ab o v e t h e h y p e rp l an e an dz e qr i f z l ie s b e l o w t h e h y p e rp l an e . S t a t is t i c a l d i s c r i m i n an t an a l y s i s i s an

es t ab l i s h ed f i e ld o f s ta t i s t ic a l t h eo ry . I n i ts s i mp l e s t , l in ea r fo rm , i t co n s i s t s i n

d e t e r m i n i n g t h e b e s t h y p e r p l a n e , i . e . t h e h y p e r p l a n e t h a t m a x i m i z e s t h e p r o b -

a b i l i ty o f c o r r e c t l y a s si g n i n g z t o p o p u l a t i o n 5 f o r qr W h e n {x i} i~ i a n d J

a r e r a n d o m s a m p l e s d r a w n f r o m m u l t iv a r i a te n o r m a l p o p u l a t i o n s w i t h re s pe c ti ve

m e a n s / ~ x a n d # r a n d common v a r i a n c e - c o v a r i a n c e m a t r i x 2~, i t c a n b e s h o w n

t h a t t h e o p t i m a l h y p e rp l an e i s g i v en b y t h e fo rm u l a (s ee F i s h e r 1 4 ]):

p t z a = O

R e s e a rc h s u p p o r t e d b y N S E R C g r a n ts 5 7 8 9 a n d 4 64 05 , th e A c a d e m i c R e s e a rc h P r o g r a m o f t h e

De par tm en t o f Na t iona l D efense (Canada) and FC A R gran t 91NC0510 . (Quebec) .

1 P ro fesso r s Dr . Pa t r i ce M arco t t e and Gi l l e s Savard , Drp ar t em en t de M ath+mat iques , Co ll~ge

mi l i t a i re roya l de Sa in t -Jean , Riche la in Qurb ec J0J 1R0 , Cana da .

03 40 - 9422 /92 /6 /51 7- 545 $2.50 9 1992 Phys i ca -Ver lag , H e ide lberg

Page 2: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 2/29

518

w h e r e

P. M arcotte and G . Savard

P = L ' - l ( # x - # r )

a = - 5 ( x + r ) r - l ( ~ x - ~ r ) - I n

a n d p a n d 1 - p a r e p r i o r p ro b a b i l it i e s a s s ig n e d t o p o p u l a t i o n s Y a n d ~1

respec t ive ly .

I n t h i s p a p e r w e f o ll o w a n a l t o g e t h e r d i f f e re n t a p p r o a c h (se e a ls o C a v a l i e r e t

a l. [3 ] ). W e d o n o t a s s u m e a n y f u n c t i o n a l f o r m f o r t h e d i s t r ib u t i o n s F x a n d F t .A c t u a l l y w e d o n o t e v e n a s s u m e t h e p o i n t s x i a n d y J t o b e r a n d o m . R a t h e r w e

l o o k f o r a d i s c r i m i n a t i n g h y p e r p l a n e t h a t m a x i m i z e s a f u n c t i o n r e p r e s e n t i n g ,

i n a s en s e to b e d e t e rm i n ed , t h e l ev e l o f d i s c r i m i n a t i o n b e t w een t h e t w o s e t s X

and Y.

T h e p a p e r i s d i v i d e d i n t o t w o s y m m e t r i c p a r t s, e a c h d e v o t e d t o a m a t h e m a t i c a l

p r o g r a m m i n g m o d e l o f t h e d i s c r im i n a t i o n p r o bl e m . T h e p a p e r is en t ir e ly d e v o t e d

t o a l g o r i t h m s a n d d o e s n o t a d d r e s s t h e i m p o r t a n t , b u t o f a v e r y d if f e re n t n a t u r e ,

q u e s t io n o f t h e r e le v a n ce o f th e m a t h e m a t i c a l p r o g r a m m i n g a p p r o a c h o v er , sa y ,

t h e s t a ti s ti c a l a p p r o a c h t o t h e d i s c r i m i n a t i o n p r o b l e m .F o r e a c h m o d e l , w e a n a l y z e w o r s t- c a s e c o m p l e x i ty , gi ve r e la t e d f o r m u l a t i o n s ,

p r o p o s e s o l u t i o n a l g o r i t h m s a n d p r e s e n t p r e l i m i n a r y c o m p u t a t i o n a l re s ul ts .

Fir s t Mode l

2 . 1 F o r m u l a t i o n

I n t h i s m o d e l , a s s u m i n g t h a t a s e p a r a t i n g h y p e r p l a n e H e x i s t s , w e s t r i v e t o

m a x i m i z e t h e m i n i m u m d i s t an ce o f an y g i v en p o i n t t o H = {z lp tz a = O, z ~ R }.

I f n o s u c h h y p e r p l a n e e x i st s, w e l o o k f o r a h y p e r p l a n e H s u c h t h a t t h e m a x i m u m

d i s t a n c e o f a n y m i s c la s si fi e d p o i n t t o t h e h y p e r p l a n e H b e m i n i m i z e d . I n b o t h

c a se s, th e h y p e r p l a n e H is d e t e r m i n e d b y s o lv i n g t h e m a t h e m a t i c a l p r o g r a m m i n gpro b le m (see Ca val i e r e t a l . I-3] ).

Page 3: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 3/29

Novel Approaches to the Discr iminat ion Problem 519

m a x w

s. t . p x i + a > w i e I = {1 . . . , r} (1 )

p y J + a < - w j e J = {1 . . . . , s }

Ilpl122 = 1

w h e r e t h e n o n l i n e a r c o n s t r a i n t IlP ll2 = 1 e n s u r e s t h a t p t x ~ + a a n d p ty j a

r e p r e s e n t t h e e u c l id i a n d i s t an c e f r o m p o i n t s x ~ a n d y J t o t h e h y p e r p l a n e H =

{ p z + a = 0 } . I g n o r i n g t h is c o n s t r a i n t i s t e m p t i n g s i n ce th e r e s u l t i n g r e l a x e dp r o g r a m i s l in e a r ( s ee F r e e d a n d G l o v e r 1 -5 ] 1 6 ], G l o r f e l d a n d G a i t h e r [ 9 ]) .

H o w e v e r it l e a d s t o in c o n s i s t e n c i e s t h a t c a n n o t b e s a t i s fa c t o r i ly r e s o l v e d , i n s p i t e

o f n u m e r o u s a t t e m p t s ( s e e K o e h l e r [ 1 6 ] , R u b i n 1 -2 1] a n d M a r k o w s k i a n d

M a r k o w s k i 1 -20]). A r e c e n t p a p e r a l o n g t h i s li n e, w h i c h r e m o v e s s o m e o f t h e

d i ff ic u lt ie s a s s o c i a t e d w i t h t h e l i n e a r p r o g r a m m i n g f o r m u l a t i o n , i s t h a t o f G l o v e r

[ 1 0 ] w h e r e t h e n o r m a l i z i n g c o n s t r a i n t

- s ~ , p x i + r ~ p y J = 1i ~ l j ~

is su b s t i t u t e d t o t h e r e v e r s e c o n v e x c o n s t r a i n t

Iip21122 ~ 1 .

I f a s e p a r a t i n g h y p e r p l a n e e x i s ts , w i s n o n n e g a t i v e a t t h e o p t i m u m , a n d ( 1) is

e q u i v a le n t t o th e re l a x e d c o n v e x p r o g r a m m i n g p r o b l e m

m a x wp a w

s . t . p~x ~ + a > w

p t y j + a < - - w

Ilp[122 < 1

i e l

j e J

(2 )

w h e r e t h e n o n l i n e a r c o n s t r a i n t i s o b v i o u s l y t ig h t a t th e o p t i m u m . W e r e fe r t o

t h i s c a s e a s t h e c o n v e x c a se . I f n o s e p a r a t i n g h y p e r p l a n e e x is ts , i.e . t h a t t h e

i n t e r s e c t i o n o f t h e c o n v e x h u l l s o f X a n d Y h a s a n o n e m p t y i n t e r i o r , t h e n ( 1 ) is

e q u i v a l e n t t o t h e r e v e rs e c o n v e x p r o g r a m :

Page 4: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 4/29

520

m a x wp a~w

s. t . p t x i + a > w

p t y j + a < - - w

IlPll 2 _ 1

i e I

j ~ J

P. Marco t te and G . Savard

(3)

o r , a f t e r h a v i n g p e r f o r m e d a n o b v i o u s c h a n g e o f v a r ia b l e :

m i n wp,a,w

s. t . p x i + a > - w i ~ I (4 )

p t y J + a < w j ~ J

I l p l l ~ > 1

w h o s e o p t i m u m w i s s t r i c t l y p o s i t i v e . T h i s l a t te r p r o g r a m b e i n g r e v e r s e c o n v e x

w e w i l l r e fe r t o t h e n o n s e p a r a b l e s i t u a t i o n a s th e r e v e r se c o n v e x c a s e .

W e w i l l a l s o r e f e r t o t h e s t r i c t l y c o n v e x c a s e i n t h e c a s e w h e r e t h e r e e x i s t s a

s t r ic t s e p a r a t i o n h y p e r p l a n e p t z + a - - 0 , i .e . t h a t w i s s t r i c t l y p o s i t i v e i n (2 ). T h e

l i n e a r c a s e w i ll c o r r e s p o n d t o t h a t s i t u a t i o n w h e r e t h e s e t s X a n d Y c a n b e

s e p a r a t e d , b u t n o t s t r ic t l y , i.e . t h a t t h e o p t i m a l v a l u e o f w i n ( 2) i s z e r o . U n d e r

o u r t e r m i n o l o g y , t h e r e i s n o s t r ic t l y r e v e r s e c o n v e x c a se .

2 .2 E q u i v a l e n t F o r m u l a t i o n

W e f ir st s h o w t h a t t h e f ir s t v e r s i o n o f t h e d i s c r i m i n a t i o n p r o b l e m D P 1 i s

e q u i v a l e n t t o s o m e d i s t a n c e p r o b l e m .

T h e o r e m 1 : T h e d i s c r i m i n a ti o n p r o b l e m D P 1 is e q u i v a l e n t t o th e m i n i m u m n o r m

p r o b l e m

m in I ]q l]2q b

s t q t x i b > 1 i ~ I

5 )

qt y J + b <_ - I j ~ J

Page 5: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 5/29

No vel Approaches to the Discrimination Problem

i n t h e s t r i c t l y c o n v e x c a se , t o t he m a x i m u m n o r m p r o b l e m

m a x Ilqll22q b

s . t . q t x i + b > - 1 i ~ l

q~yJ + b < l j e J

i n t h e r e v er s e c o n v e x c a se , a n d t o th e m a x i m u m n o r m p r o b l e m

m ax I lq ll2q b

s t qtxi 4- b >_ i ~ I

q t yJ + b _ < O j e J

- l _ < q k _ < l k = l . . . . . n

i n t h e l i n e a r c a s e .

521

(6)

( 7 )

Proof.

(a) S t r i c t l y c o n v e x c a s e . F i r s t n o t i c e t h a t a n y o p t i m a l s o l u t i o n t o (2 ) m u s t s a t is f y

p 0 . O t h e r w i s e w e w o u l d h a v e a > w a n d a _< - w , i.e . w = 0 , a n i m p o s s i b i l i t y

s i n c e w > 0 a t t h e o p t i m u m i n t h e s t r i c t ly c o n v e x c a s e . S i m i l a r l y , q 0 a t th e

o p t i m u m o f (5 ).

L e t ( p *, a * , w * ) b e a n o p t i m a l v e c t o r f o r (2 ). T h e n ( q* , b * ) = ( p * / w * , a * / w * ) is

w e l l d e f i n e d f o r ( 5 ) . A s s u m e i t i s n o t o p t i m a l f o r ( 5 ) , a n d l e t ( ~ , / ~ ) b e o p t i m a l

ins tea d . T h e n I1 11 < I Iq*l l a n d ( if , a , i f ) = ( / 1 1 ~ 1 1 , b / l l l l , 1 / 1 1 1 1 ) i s f e a s i b l e f o r ( 2 )

w i t h f f = 1 / 1 1 ~ t l > 1/llq*ll = w / l l p l l = w * ( t h e c o n s t r a i n t I I p l l < 1 i s t i g h t a t

t h e o p t i m u m ) , i n c o n t r a d i c t i o n w i t h t h e o p t i m a l i t y o f w * . H e n c e ( q* , b * ) i s

o p t i m a l f o r (5 ).

N o w l e t ( q *, b * ) b e o p t i m a l f o r (5 ) a n d d e f i n e (p * , a * , w * ) = ( q * / l l q * 1 1 , a / l l q I I ,1 /llq *ll). A g a i n , a s s u m e t h a t ( p * , a * , w * ) i s s u b o p t i m a l , l e t (/~ , a , f ) b e o p t i m a l

w i t h f > w * a n d d e f i n e ( iT , b ) = ( / ~ /f , a / f ) . W e h a v e : 1 1 4 1 1= I l P l l / f = l i f t <

I / w * = IIq * II, i n c o n t r a d i c t i o n w i t h t h e o p t i m a l i t y o f q * .

(b) R e v e r s e c o n v e x c a s e . F i r s t, n o t e t h a t , i n th e r e v e r s e c o n v e x c a s e , p r o b l e m (6 )

is b o u n d e d . O t h e r w i s e l e t (q(k), b(k)) b e a fe a si bl e, u n b o u n d e d s e q u e n ce . W e h a v e :

q k ) t b k ) - - 1

x i + >tl(q k), btkJ)l II(q k), btk))lr - - iI(qtk) , btk~)l

L e t (q *, b * ) b e a n y c o n v e r g e n t s u b s e q u e n c e o f t h e b o u n d e d s e q u e n c e

Page 6: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 6/29

522

(q(k)/ll(q (k), b(k))ll, b(k)/ll(qtk) b(k))ll) .

W e t h e n h a v e , b y c o n t in u i t y ,

q t x i + b * > 0

P. M arcotte and G. S avard

an d , s i mi l a rl y :

q , t y j + b * <_ 0

an d (q * , b * ) d e t e rmi n es a (n o t n eces s a r i l y s t r i c t ) s ep a ra t i o n h y p e rp l an e , an i m-

p o s s i b i l i t y s i n ce w e a r e i n t h e r ev e r s e co n v ex ca s e . N o w , u s i n g t h e o n e - t o -o n e

re l a t i o n s h i p b e t w een f ea s ib l e s o l u t i o n s t o (3) (w i t h w > 0 ) an d f ea s i b le s o l u t i o n s

t o 6 ) :

(q, b) = ( p / w , a / w ) ,

t h e p r o o f i s s i m i la r t o t h e p r e v i o u s o n e , m o d u l o s i gn r e ve r sa l.(c) L i n e a r c a s e . A ny feas ib le so lu t ion to (7) wi l l p ro v ide a feas ib le so lu t io n , i .e . a

s o l u t i o n w i t h a n o n z e r o v e c t o r p . [ ]

R e m a r k : In t h e s t r i c t l y co n v ex ca s e , t h e d i s c r i m i n a t i o n p ro b l em i s eq u i v a l en t t o

p r o j e c ti n g t h e o r i g i n o n t h e c lo s e d , c o n v e x p o l y h e d r o n

{ q e R n l q t x + b > l V i e I

q t y J + b < - i V j ~ J

fo r s o m e b ~ R} .

T h i s p o l y h e d ro n d o es n o t co n t a i n t h e o r i g i n (s ee f ig u re 1 ).

I n t h e r ev e r s e co n v ex ca se , i t r ed u ces t o f i n d i n g t h e ( ex tr eme) p o i n t o f m ax i m u m

e u c l id i a n n o r m o f th e c lo s ed c o n v e x p o l y h e d r o n

{ q E R n [ q t x i + b > - 1 V i ~

q ty J + b < l V j ~ J

fo r s o m e b ~ R}

Page 7: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 7/29

P 2

N o v e l A p p r o a c h e s t o t h e D i s c r i m i n a t i o n P r o b l e m 5 23

O

F i g 1 : T h e c o n v e x c a s e

I J

P l

h a v i n g t h e o r i g i n a s a n i n t e r i o r p o i n t ( se e f i g u r e 2 ). I n t h e l i n e a r c a s e , p r o b l e m

(5 ) is in f e a si b le , w h i le p r o b l e m (6 ) is u n b o u n d e d .I t is a l s o p o s s i b l e to r e f o r m u l a t e t h e d i s c r i m i n a t i o n p r o b l e m , i n t h e r e v e r s e

c o n v e x c a s e , a s a b i l ev e l p r o g r a m m i n g p r o b l e m .

Theorem 2: I n t h e r e v e r s e c o n v e x c a s e , (3 ) a n d (6 ) a r e e q u i v a l e n t t o t h e l i n e a r

b i l e v e l p r o g r a m :

~ 1 t . e 2 t u 2m a x c u +

p a

s.t . A p + ae I >_ - e 1

Bp + ae 2 <_ e 2

( u l , u z ) 6 a r g m a x -e l~ u 1 e2 u 2 (8)

s . t . - A t u 1 + B~u 2 = p

e l t u I J e 2 t u 2 = 0

u 1, u 2 >_ 0

w h e r e A = ( x 1 . . . , x ) t, B = ( y l . . . . . y~)r, e 1 = ( 1 . . . . . 1 )r e R r a n d e 2 =

1 . . . . . 1 ) t ~ R s .

Page 8: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 8/29

P2

F i g 2 : T h e r e v e r se c o n v e x c a s e

5 2 4 P M a r c o t t e a n d G S a v a r d

Pl

Proof The transformation of a concave quadrat ic program into a linear bilevel

program has been shown to be valid in Konno [181 using linear dualityarguments. []

2.3 Com plexity Ana lysis

Theorem 3: The discrimination problem DP1 is in P in the convex (strictly convex

9 r linear) case and strongly NP-complete in the reverse convex case.

Proof (a) In the strictly convex case, the result follows from the existence of

polynomial algorithms for the monotone linear complementari ty problem which

constitutes an extension of the convex quadratic programming problem (see

Kojima et al. [14]).

(b) In the linear case, there must exist a nonzero vector p and a scalar a such

that the set of linear inequalities

p t x i + a > _ O i ~ I

p t y ~ + a < 0 j ~ J

is satisfied. One way to enforce the condition p ~ 0 is simply to impose the

Page 9: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 9/29

Novel Approaches to the DiscriminationProblem 525

a d d i t i o n a l c o n d i t i o n Pk > 1 or Pk < -- 1) f o r s o m e i n d e x k . S in c e it is u n k n o w n

a priori w h i c h i n d e x k w i ll c o r r e s p o n d t o a n o n z e r o c o e t ti e ie n t P k, o n e h a s t o

s o lv e , in t h e w o r s t c a s e , t h e 2 n l i n ea r f ea s i b i l it y p ro b l em s :

p t x i + a > _ O i ~ I

pty~ + a < 0 j ~ J 9a)

pk > 1

a n d

p t x i + a > O i e I

pty~ + a < 0 j e J 9b)

Pk <-~ - -1

f o r k = 1 . . . , n , a n d p o l y n o m i a l i t y i s a c o n s e q u e n c e o f t h e p o l y n o m i a l i t y o f l i n e a rp r o g r a m m i n g see K h a c h y a n [1 5 ]) .

c ) I n t h e r e v e rs e c o n v e x ca s e, th e s t r o n g N P - c o m p l e t e n e s s o f D P 1 f o l lo w s

f r om t h e s tr o n g N P - c o m p l e t e n e s s o f t h e m a x i m u m n o r m p r o b le m se e F r e u n d

a n d O d i n [ 7 ] o r B o d l a e n d e r e t a l. [ 2 ]) . [ ]

Coro l lary: T h e e x i s te n c e o r n o n - e x i s t e n c e o f a d i s c r i m i n a t i n g h y p e r p l a n e c a n b e

d e t e r m i n e d i n p o l y n o m i a l ti m e .

Proof. F ro m t h eo r em 3 , p a r t b ), t h e r e fo l lo w s t h a t t h e exi s ten ce o f a d i s c r i mi n a t i n g

h y p e r p l a n e c a n b e d e t e r m i n e d b y s o l v in g fo r t h e e x i s te n c e o f a f e as ib l e s o l u t i o n

t o o n e o f 2 n l i n ea r s y s t ems . H en c e t h e r e s u lt . [ ]

2 .4 S o lu t i o n A lg o r i t h m s a n d N u m er i ca l R es u l t s

So l u t i o n a l g o r i t h ms a r e b a s ed o n t h eo rem 1, w h i ch s t a t e s t h e eq u i v a l en ce b e t w een

D P 1 a n d q u a d r a t i c p r o g r a m m i n g . I n t h e c o n v e x c a se i t i s e a s il y s o l v e d a s a

p r o j e c t i o n p r o b l e m f o r w h i c h e ff ic ie n t a l g o r i t h m s e x is t se e W o l f e [ 23 ] o r t h e

s u r v e y b y L i n a n d P a n g [ 1 9 ]) . I n t h e r e v e r se c o n v e x ca s e, w e s o l v e d D P 1 u s i n g

a c o d e i n i ti a ll y d e s i g n e d f o r so l v i n g l i n e a r bi le v e l p r o g r a m m i n g p r o b l e m s se e

Page 10: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 10/29

526

Table 1: Possible param eters combinations

P. M arcotte and G . Savard

#x 0.0 0.0 0.0 0.0 0.0 0 .0

Dispersion # r 0.5 1.0 2.0 4.0 8.0 i6.0parameters Ex I I I I I IZr I I 41 41 161 161

Sam plesize (r,s) (20,10) (20,15) (20,20) (30,15) (30,22) (30,30) (40,20) (40,30) (40,40)

Dimension n 2 3 4

H a n s e n e t a l [ 11 ]) . T h i s a l g o r i t h m i m p l e m e n t s i d e as o f l o gi c p r o g r a m m i n g w i t h i n

a B r a n c h - a n d - B o u n d s c h em e a n d c o n v e rg e s t o a g l o b l o p t i m u m i n f in i te l y m a n y

s te p s. N e c e s s a r y o p t i m a l i t y c o n d i t i o n s e x p r e s s e d i n t e r m s o f t i g h t n e s s o f t h e

s e c o n d l e v e l ' s c o n s t r a i n t s a r e u s e d f o r f a t h o m i n g o r s i m p l i f y i n g s u b p r o b l e m s

t h r o u g h e l i m i n a t i o n o f v a r i a b l e s, b r a n c h i n g , a n d o b t a i n i n g p e n a l t i e s si m i l a r to

t h o s e u s e d i n m i x e d i n t e g e r p r o g r a m m i n g . F o r t h o s e p r e l i m i n a r y r e s u l t s , n o

a t t e m p t s h a v e b e e n m a d e a t s p e c ia l iz i ng th i s a l g o r i t h m t o t h e p a r t i c u l a r s t r u c t u r e

o f 8 ) .O u r e x p e r i m e n t a l d e s i gn f o l lo w s c lo s e ly t h a t o f B a jg i er a n d H i l l [ 1 ]. T h e

d i s c r i m i n a t i n g p o i n t s a r e d r a w n f r o m m u l t i v a r i a t e n o r m a l p o p u l a t i o n s w i t h

v a r i a b l e d i s p e r s i o n p a ram e t e r s (# x , / ~ r , 2 7x , 2 7 r ), s am p l e s i ze s r , s an d s p ace

d i m e n s i o n n . T a b l e 1 s h o w s t h e c o m b i n a t i o n s u s e d f o r t h e a c t u a l e x p e r i m e n t s .

F o r e a c h o n e o f t h e 6 x 9 x 3 = 1 62 c o m b i n a t i o n s t h a t h a v e b e e n t e s t e d , a s e t

o f 1 0 r a n d o m p r o b l e m s h a s b e e n g e n e r a t e d . N u m e r i c a l r e s u lt s a r e p r e s e n t e d i n

t a bl e s 2, 3 a n d 4 , w h e r e m r e f er s t o t h e a v e r a g e ( C P U - t i m e a n d n u m b e r o f n o d e s

e x p l o r e d i n th e b r a n c h - a n d - b o u n d s c h em e ) o f t h e c o r r e s p o n d i n g s e t o f p r o b l e m s ,

s , , t o t h e s t a n d a r d d e v i a t i o n a n d m e t o t h e m e d i a n o f t h e o b s e r v a t io n s , T h e

b il ev e l p r o g r a m m i n g c o d e, w r i tt e n in F o r t r a n , h a s b e e n r u n o n a S U N S P A R K

w o r k s t a t i o n .

A s e x p e c t e d , th e p r o b l e m ' s d i f f ic u l ty i s c l o s e ly r e l a t e d t o t h e n u m b e r o f n o d e s

e x p l o r e d in t h e b r a n c h - a n d - b o u n d s c h em e u n d e r l y i n g t he b i le v el p r o g r a m m i n g

a l g o r i th m , a n d i n c re a s e s w i t h t h e d e g r e e o f o v e r l a p b e t w e e n t h e t w o s et s o f p o i n t s

X a n d Y. W h e n t h e o v e r l a p i s w e a k , t h e n t h e n u m b e r o f n o d e s t o b e e x p l o r e d in

o r d e r t o o b t a i n a n o p t i m a l s o l u t i o n is a l m o s t i n s e ns it iv e t o t h e s p a c e d i m e n s i o n

n. T h i s h o w e v e r is n o t t h e c a s e w h e n t h e o v e r l a p i s s t r o n g , i.e . w h e n t h e m e a n s

# x an d # r a r e c l o s e . F o r i n s t an ce , w i t h t h e d a t a s e ts (# x , # r , 2 7x , Z ' r ) = (0, .5 .; I , I )

a n d ( # x , # r , S x , 2 ; r ) = (0 , 1, I , I ) t h e n u m b e r o f n o d e s e x p l o r e d i s a p p r o x i m a t e l y

t h e s a m e i n d i m e n s i o n 2 ( n = 2 ) w h i l e i t i s i n th e r a t i o 2 to 1 w h e n t h e d i m e n s i o n

i s i n c r e a s e d f r o m n = 2 t o n = 4 . W e a l s o n o t e d t h a t t h e m o s t d i ff ic u l t p r o b l e m s

u s u a l ly c o r r e s p o n d e d t o t h e s i t u a t io n w h e r e t h e n u m b e r o f p o i n t s i n e a c h s a m p l e

we re c l o s e t o eac h o t h e r ( r ~ s ). Fo r t h e ea s i e r p ro b l em s ( r << s o r s << r ) t h e

m e d i a n w a s c l o s e t o t h e a v e r a g e a n d t h e s t a n d a r d d e v i a t i o n w a s s m a ll , a n

u n u s u a l f e a t u r e i n t h e s o lu t i o n o f p r o g r a m s v i a b r a n c h - a n d - b o u n d m e t h o d s .

Page 11: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 11/29

T

e2

T

b

e

a

o

oD

P

d

as

# /Z E En

r=

2

B

sm

s=

1

me

r=

2

B

s=

Sm 

me m 

r=

2

B

sm 

s=

2

me

r=

3

B

s=

i  m 

me m 

r=

3

B

Sm 

s=

2

m

e

r=

3

B

s=

3

r=

4

B

s=

2

r=

4

B

s=

3

r=

4

B

s=4

m  m me m  Sm

 me m  Sm

 me m  m  me

00

05 I I 2

00

80 I

12

00

1

0I

12

n

c

n

c

n

n

c

4

8

3

5

1

6

31

3

3

3

5

6

6

1

1

8

1

6

5

6

8

5

6

8

0

1

8

2

4

5

8

0

7

4

1

3

2

9

3

4

7

1

1

1

4

2

0

2

1

4

8

1

2

5

1

8

4

3

6

1

1

3

1

3

8

1

6

3

2

3

0

9

3

1

3

0

4

2

1

0

3 5

8

1

5

5 6

4

1

3

6 7

8

2

3

71

4

3

7

1 1

2

4

0

1 1

6

4

0

1 2

4

9

3

1 2

0

1

1

2

00

00

10

20

  I

4

2

2

c

n

3

3

2

2

1

7

6

3

3

2

5

6

3

6

1

0

8

7

6

0

3

9

0

5

8

1

7

1

1

8

0

4

1

9

5

8

3

2

1

2

1

0

4

2

3

8

6

7

4

2

8

2

7

7

4

7

8

0

1

4

1

8

3

2

8

3

9

9

4

1

1

2

7

3

4

9

7

1

1

2

3

1

3

8

6

0

1

1

3

1

2

8

3

3

3

9

2

1

2

367

2

3

4

4

1

6

4

8

7

2

2

0

7

0

8

0

2

9

8

5

1

6

5

1

1

5

2

1

6

8

2

1

2

6

8

3

3

4

5

6

1

4

4

9

8

6

2

5

7

4

2

8

7

5

2 3

8

9

0

3 3

2

1

8

3 3

8

t1

3

3 7

0

1

4

6 7

0

3

5

8 5

6

2

5

5 113o 8 

2

5

1 1

8

6

7

1

00

40 I 4 2

c

n

2

3

2

2

79

4

1

2

8

2

4

0

2

6

1

4

7

0

3

5

2

5

1

3

0

1

5

1

8

5

5

3

7

6

4

8

2

8

1

6

7

2

3

1

9

4

6

4

3

2

0

1

8

3

2

5

4

4

1

5

2

3

2

0

5

1

5

6

2

7

8

2

9

1

4

6

4

2

7

8

1

2

3

1

4

9

7

8

6

6

8

4

1

2

7

6

4

7

1

0

6

9

6

6

2

0

1

5

0

1

1

4

6

9

8

1

6

9

5

2

2

2

1

6

2

1

0

3

3

1

8

3

4

5

2

2

3

5

5

c 2

4

1

6

39

6

0

2

2

1

3

1

2

4

89

7

9

2

4

2

4

7

2

0

2

6

7

5

5

3

2

7

6

2

8

2

0

1

3

7

0

2 4

0

1

5

4 5

2

1

4

5 4

4

2

2

5 6

6

1

3

6 7

0

2

6

7

2

4

1

1

6

3

3

8

2

1

7

1

3

56

 48

2

888

2

5

4

7

1

5

4

8

5

7

2

6

4

9

1

1

2

8

1

6

1

4

4

5

1

3

2

0

8

4

1

8

2

9

7

3

3

2

4

7

1

3

4

4

ZO ;> , Oc  4

Page 12: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 12/29

T

e3

T

b

e

a

o

oD

P

s

d

as

r=

2

s=

1

r=

2

s=

r= 

2

s=

2

r=

3

s=1

r=

3

s=

2

r=

3

s=

3

r=

s=

2

r

=

4

s=

3

r=

4

s=

4

00

10

I I 3

00

20

I 4 3

00

40 1 4 3

00

80 I

13

00

1

0I

13

c

n

c

n

c

n

c

n

c

n

c

#

00

#

05

Ex

1

E

I

n

3

n

6

2

sm

1

4

me

6

9

4

sm

2

1

me

1

1

8

sm

2

7

me

1

1

0

s~

2

9

me

1

2

4

s

6

8

me

2

2

4

s

7

6

me

2

2

8

sm

8

0

me

2

3

4

sm

8

6

me

3

3

8

s=

6

0

me

3

6

7

1

6

6

4

5

8

4

3

1

9

1

5

5

4

5

1

2

4

8

1

9

4

2

4

4

1

2

1

0

4

4

1

B

1

9

7

2

9

9

6

2

8

2

B

3

0

2

6

3

4

2

6

2

6

1

9

7

9

8

5

8

0

B

2

5

5

9

2

8

4

6

1

9

4

6

8

4

2

9

7 8

3

2

0

6

3

B B

1

2

1

3

2

2

3

5

9

1

5

1

6

1

6

4

9

7

6

8

1

4

1

2

4

4

7

0

1

5

1

3

6

2

6

7

3

6

8

2

5

2

6

B

2

6

6

9

7

4

2

1

2

7

3

3

9

1

5

1

3

5

5

3

1

1

2

9

1

7

1

7

8

7

2

1

6

B B B

8

0

1

5

2

7

4

8

7

1

3

9

8

1

0

3

4

5

8

8

1

1

1

6

3

3

5

6

1

8

1

2

4

1

8

5

3

3

0

1

5

1

5

5

1

9

3

5

4

1

1

5

1

4

8

2

6

8

8

9

7

3

7

2

8

3

2

4

 9

7

7

2

1

9

4

7

1

5

B

1

4

1

5

1

1

1

8

8

79

2

3

6

1

8

2

3

2

3

7

2

6

2

7

1

3

1

1

2

2

3

3

4

4

8

1

3

1

5

3

4

4

9

2

96

1

4

1

1

3

45

1

2

3

3

1

3

1

6

2

6

1

2

1

4

9

1

9

4

4

7

8

7

4

1

3

2

2

4

6

2

1

4

4

1

4

3

6

1

5

2

9

4

7

1

"6

1

9

1

3

5

4

4

8

4

5

0

1

7

1

8

3

6

2

2

4

9

2

2

4

7

1

8

2

6

3

6

8

5

9

6

1

3

1

7

6

1

0

9

4

2

1

3

0

1

2

3

4

9

7

3

8

6

7

2

6

7

0

7

1

7

8

2

1

2

6

7

9

6

3

5

1

8

3

1

7

2

2

9

4

2

1

5

2

5

7

5

3

7

1

4

8

2

A

1

3

3

5

2

6

7

6

2

3

8

6

2

1

2

6

9

8

4

6

1

7

4

0

1

2

2

7

7

4

4

1

1

8

2

6

1

6

2

1

0

6

1

3

2

1

1

5

9

1

6

1

6

8

3

5

A

1

1

9

1

1

2

9

1

6

1

0

5

8

5

1

1

4

1

4

3

1

8

8

3

6

9

3

9

1

8

8

9

4

4

5

2

0

1

6

9

4

7

8

8

5

9

5

6

3

8

8

4

4

o~ 

c o~ 

Page 13: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 13/29

Ta

e4

Th

b

e

a

o

oD

PI

h

d

d

a

s

fx

la  n m 

r=

2

BLP 

s

s

1

me

r=

2

BLP 

s=1

r= 2

BLP 

s=

2

r= 3

BLP 

s=1

In Sm 

me m

 Sm 

me

Sm 

me m 

r=

3

BLP 

sm 

s=

2

me m 

r=

3

BLP 

sm 

s=

3

me

r= 4

BLP 

s=

2

r= 4

BLP 

r=

3

r=

4

BLP 

s=

4

Sm 

me m

  m 

me m

 Sm 

me

0

0

0

5I I 4

0

010

1 I 4

0

0

2

0

I 4 4

0

0

4

0I 4  

0

08

14

n

c

n

c

n

c

n

e

n

c

9

6

1

5

3

8

3

2

9

9

5

1

0

2

7

6

2

8

3

1

2

4

2

2

3

3

6

5

1

3

1

2

3

2

8

5

1

7

7

1

9

2

4

9

4

4

1

0

4

1

0

3

2

1

4

2

3

5

8

2

2

1

9

4

9

1

7

3

2

5

1

2

0

4

2

1

3

4

7

6

0

1

3

6

4

2

5

6

3

7

4

3

2

3

5

9

6

1

9

4

3

5

0

8

1

9

1

2

1

8

3

0

6

0

1

1

7

1

2

2

3

7

6

1

2

1

2

6

2

4

9

5

1 3

2

1

3

3 3

2

8

1

2 6

8

2

4

5 6

4

2

3

4

5

2

2

6

5

2

1

3

3

1

1

7

9

2

2

1

8

7

2

5

8

3

2

6

3

8

1

7

2

2

4

8

5

0

2

1

2

8

3

4

9

9

2

1

6

2

7

3

3

8

1

3

1

0

2

7

4

2

7

0

1

2

0

1

6

2

9

3

9

6

3

1

1

0

828

9

6

9

0

1

0

3

51

1

0

2

5

2

7

3

4

4

75

2

6

4

2

2

0

3

6

1

3

8

5

8

1

1

5 7

7

1

7

7

1

6

3

8

1

1

3

4

1

2

1

2

8

1

3

2 5

2

3

3

3 9

6

3

7

9 4

0

3

4

3

7

9

1

319

1

9

2

38

7

2

6

2

4

1

7

2

4

2

8

1

2

2

0

5

3

5

6

1

7

1

6

6

5

8

0

2

3

5

4

1

6

7

9

6

5

1

4

0

2

6

3

9

2

2

2

8

9

0

9

0

3

4

1

6

1

7

1

8

3

7

7

5

2

7

7

5

1

2

2

9

2

7

2

2

1

1

5

2

2

8

5

6

2

2

5

1

4

8

8

9

6

3

0

7

3

2

0

5

2

1

0

2

8

4

4

1

0

1

3

9

4

2

5

I 838 

0

0

1

01

14

c

n

2

6

6

4

1

9

1

9

2

1

1

7

5

1

0

4

0

1

4

4

0

6

1

0

1

0

5

2

1

2

1

5

5

1

4

1

4

7

6

1

3

8

4

4

2

7

2

0

1

3

2

7

2

8

6

6

3

4

8

2

4

5

3

6

5

2

4

7

2

2

1

7

3

2

4

5

6

1

9

5

0

4

0

5

4

1

6

4

1

4

4

6

9

6

6

4

1

3

1

4

0

1

9

8

7

3

5

4

1

8

5

6

3

2

1

1

8

1

5

3

4

1

4

5

4

1

3

4

1

6

2

6

2

4

1

3

2

1

2

1

2

[

2

2

1

3

7

9

I

1

5

1

605 6 

I 1

2

4

2

2

2

7

2

4

3

0

7

2

Z0 1 0 l

Page 14: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 14/29

53

3 S e c o n d M o d e l

P. M arcotte and G. S avard

3 . 1 F o r m u l a t i o n s

M i n i m i z i n g t h e n u m b e r o f m i c la s si fi e d p o i n t s r e s ul ts i n t h e m i x e d i n t e g e r p r o -

g r a m m i n g p ro b l e m D P 2 :

m i n n x + n rp a

w h e r e n x = c a r d { x ~ l p t x ~ + a < 0 } an d n r = c a r d { y J [ p t y j + a > 0 } . I n t r o d u c i n g

b i n a ry v a r iab l e s ~ (~ = 0 i f x i i s no t m isc lassi f i ed , 1 o therw ise) an d t/ j (qj = 0 i f

y J is n o t mi s c l a s si f ied , 1 o t h e rw i se ) , an d o b s e rv i n g t h a t w e m u s t i m p o s e in s o m e

w a y th e c o n s t r a i n t p 0 , w e o b t a i n t h a t D P 2 c a n be s o lv e d b y s o l v in g t h e t w o

m i x e d i n t eg e r p r o g r a m s

m i n ~ ~ + ~ t / jp a ~ ~l i~ l jE J

s.t. M ~ i >_ - p t x i + a ) i ~ I

M t l j >_ p t y j + a J ~ J ( l l a )

~ j e { o 1}

- - l < p k < _ l k = l . . . . , n

P l = - 3

a n d

m i n 2 ~ , + 2 r tJp a ~ ~l i~ l j~ J

s.t. M~i >_ - - p tx i + a) i e I

M t l j > p t y ~ + a J ~ J (1 lb)

~i, ~/j ~ {0, 1}

- - l < _ p k < _ l k - - 1 . . . . . n

P i = 6

Page 15: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 15/29

N o v e l A p p r o a c h e s t o t h e D i s c r i m i n a t i o n P r o b l e m 5 3

w h e r e M is a p o s i t iv e c o n s t a n t s u f f ic i e nt ly l a r g e t o e n s u r e t h a t p r o b l e m s l l a )

a n d 1 l b ) a r e f e a s ib l e f o r a ll a d m i s s i b le v a l u e s o f p a n d a a n d 6 i s a s m a l l p o s it i v e

c o n s t a n t . W e h a v e t h e f o ll o w i n g t h e o r e t i c a l r e s u lt :

T h e o r e m 4 : F o r D P 2 t o b e e q u i v a l e n t t o l l a ) a n d l l b ) i t is s u ff ic i en t t h a t th e

c o n d i t i o n

M > 2 t m a x ~ xgl m a x ~ tY ~ t}i~l k = l j ~ J k = l

b e s a t i s f ie d .

Proof . I f a < - m a x / ~ k I x~ l < ~ k p k X ~ f o r a l l p o i n t s x ' i n x , r e c al l t h a t [PRI < 1)

t h e n a l l p o i n t s i n X a r e m i s c la s s i fi e d . O n t h e o t h e r h a n d , i f a > m a x ~ ~ k l x ~ l , t h e n

a l l p o i n t s i n X a r e w e l l c l a s si f ie d . A s i m i l a r r e s u l t h o l d s f o r p o i n t s i n Y , r e p l a c i n g

x i b y y J i n th e a b o v e i n e q u a li ti e s . W e c a n t h e r e f o r e c o n c l u d e t h a t , w i t h o u t l o s s

o f g e n er al it y , la l c a n b e b o u n d e d b y m a x , ~ , ~ = l l x ~ l + m a x i ~ s ~ = x l Y ~ l :

C o n s e q u e n t l y , I p t x i + a l a n d I p ' y j + a l w i l l a l w a y s a s s u m e v a l u e s l e s s t h a n2 { m a x , ~ , ~ = ~ Ix~ l + m a x j ~ s ~ = ~ lY/~I}, a s c l ai m e d . [ ]

O n e c a n s u b s t i t u te t o th e m i x e d i n te g e r p r o g r a m s 1 l a ) a n d l l b ) t h e s in g le

b i le v e l p r o g r a m :

m i n ~ , + ~ q j + L [ ~ i + X [ 3 i + 7 1p,a,~,~,lt i~l j~ J L eI je J

s.t. M ~ i > - - ( p t x i + a )

M r l j > p t y j + a

2/~ + p l = 3

- - l < p k < l

i ~ I

j ~ J

k = 1 , . . . , n

~ , f l , ? ) ~ a r g m a x ~ o q + ~ f l j +

i ~ l j ~ J

s.t. ~i -< r ~i < 1 - ~i

r > 0 i ~ I ,

,8j <_ rlj ,Sj <_ l - rli

1 2 )

Page 16: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 16/29

532 P. M arcotte and G. Savard

t l j > O j e J

? _ < p ? < 1 - / ~

? _ > 0

w h e re L i s a s u i t ab l y l a rg e co n s t an t . T h e c o n s t r a i n t s e t o f th e l o w e r lev e l f o r ce s

v a r ia b l e s ~ , f lj a n d ? t o t a k e n o n z e r o v a l u e s w h e n e v e r a b i n a r y v a r i a b l e ~ , r/i

o r is n o t e q u a l to z e r o o r o n e . I n o r d e r t o m i n i m i z e t h e s u m o f t h e ~ ' s , fi rs

an d ? , t h e u p p e r l ev el w i ll s et t h e co n t i n u o u s v a r i ab l e s (~ i's , ~ / fs an d / ~ ) , t o e i t h e r

0 o r 1. I n d e e d , t h e l o w e r l e v el s i m p l y t r a n s la t e s i n t o a n o p t i m i z a t i o n f o r m t h e

re l a t i o n s h i p s

~i = m in{ ~i, 1 - ~i}

fli = min{ ~/j, 1 - ~/j}

? = m i n { , 1 - / ~ } .

T h e l o w e r le v el p r o g r a m , t o g e t h e r w i t h t h e u p p e r l ev e l c o n s t r a i n t

2/~ + p l = 1

e n s u r e s t h a t P l w i ll t a k e e it h e r v a l u e - 1 o r 1, t h u s a v o i d i n g t h e d e g e n e r a t e

s o l u ti o n . T h e s o l u t i o n t o D P 2 w i ll n o t i n g e n e r a l be u n i q u e , a n d s o l u t i o n m e t h o d s

b a s e d o n l i n e a r p r o g r a m m i n g r e l a x a t i o n ( b r a n c h - a n d - b o u n d , c u t t i n g p l a n e s ,

L ag ran g ean r e l ax a t i o n ) w i l l f av o r ex t r ema l s o l u t i o n s i . e . , i n t h i s co n t ex t ,

h y p e r p l a n es g o i n g t h r o u g h a t l e a s t n p o i n t s o f e i t h e r X o r Y (s ee f i g u re 3 fo r a

t w o - d i m e n s i o n a l e xa m p l e ). I n t h i s c a s e it is n a t u r a l t o s e e k, a m o n g t h e s e t o f

h y p e r p l a n e s m i n i m i z i n g t h e n u m b e r o f m i s c la s si fi e d p o i n ts , a n h y p e r p l a n e t h a t

mi n i m i zes a s w e l l t h e s e t o f w e ll -c l a ss if i ed p o i n t s . T h u s w e o b t a i n t h e m e t a -b i l ev e l

f o r m u l a t i o n :

m a x m i n ( m in f x i a,p a Ix i r welliX)

s .t . (p , a) ~ ar g m in n x + nr

m i n _ p t y J _ a }yJ r well Y)

1 3 )

IlPll2 ~ 1

w h e re w e l l (X ) an d w e l l(Y ) d e n o t e t h e s e ts o f w e l l -c l a ss i fi ed p o i n t s i n X an d Y

respec t ive ly , whi l e n x ( r e s p ect i v e ly n y ) d en o t e s t h e n u m b e r o f mi s c l a s s if i ed p o i n t s

in X (resp ect ively in Y).

Page 17: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 17/29

Novel Approaches to the DiscriminationProblem 533

P~ 0

..o%o0 0 O

9 ~ a = O

m

P l

Fig. 3: An exterm al solution

In v i ew o f t h e d i f f i cu lt y o f s o l v i n g (1 3 ) , w e w i ll in s t ead co n s i d e r a g o a l p ro -

g r a m m i n g f o r m u l a t i o n w i t h p r i m a r y o b je c ti ve t h e n u m b e r o f m i sc l as s if ic a ti o nsan d s eco n d a ry o b j ec t iv e t h e s ep a ra t i o n o f t h e se t s o f w e l l -c l a ss if i ed p o i n t s . T h i s

r e s u l t s in t h e fo l l o w i n g p ro ced u re :

G o a l P r o g r a m m i n g A l g o ri th m

1. L et (p*, a* ) e ar g m inp, a n x + n r a n d I x ( r e s p ec ti v e l y Iv ) d en o t e t h e i n d ex s e t

o f wel l -c lass i f ied po in t in X ( respec t ive ly Y).

2 . So l v e t h e p ro j ec t i o n p ro b l em :

min t lql l2q b

s.t. q tx l + b >_ 1 i E I x

q t y J + b < - I j e l r 9

(14)

I f (14) i s feas ib le ( s t r i c tly con vex case) l e t (q* , b*) be i t s so lu t io n an d se t: p* =

q* /tlq * II 2, a * = a*/llq*}l 2.3 . O u tp u t (p* , a* ) . [ ]

Page 18: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 18/29

534

3 .2 Com plex i ty Ana l ys i s

P. M arcotte and G . Savard

I n t h e c o n v e x c a se , t h e o p t i m u m o f (14) is z e r o , a n d o u r g o a l p r o g r a m m i n g

f o r m u l a t i o n r e d u c e s D P 2 t o D P 1 . I n t h e r e v e r s e c o n v e x ( o r i n t h e g e ne r a l) ca s e,

w e o b t a i n t h e f o l l o w i n g w o r s t - c a s e r e s u lt :

T heorem 4 : T h e d e c i s i o n p r o b l e m c o r r e s p o n d i n g t o D P 2 i s s t r o n g l y N P -

c o m p l e t e .

Proof T h e d e c i s io n p r o b l e m c o r r e s p o n d i n g t o D P 2 c a n b e s t a te d a s:

D o e s t h e r e e x is t a h y p e r p l a n e H s u c h t h a t t h e n u m b e r o f m i s c l a s si f ic a t io n s

i s l es s o r e q u a l t o K ?

(a ) D P 2 i s i n N P .

L e t I d e n o t e t h e c a r d i n a l i t y o f a m a x i m a l s e t S o f a f f in e l y i n d e p e n d e n t p o i n t s i n

X w Y an d l e t n ' = ca rd (S ) = r a in { l , n }. In p rac t i ce we wi l l u s u a l l y h av e t h a t r + s

i s m u c h l a r g e r t h a n n a n d w e e x p e c t n ' = n . T h e r e a l w a y s e x i s ts a s o l u t i o n ( p, a )t o D P 2 s u c h t h a t t h e n ' p o in t s i n S li e o n t h e h y p e r p l a n e ptz + a = 0 . T h e re fo r e

a s o l u t i o n t o D P 2 c a n b e c h a r a c t e r i z e d b y s p e c i f y in g t h e f i n it e s e t S . F o r a g i v e n

S , t h e v a r i a b l e s p a n d a a r e d e t e r m i n e d a s s o l u t i o n s o f t h e ( p o s s ib l y u n d e r -

d e t e r m i n e d i f n ' < n ) l i n e a r s y s t e m

tz + a = 0 fo r al l z e S ,

a n d , p r o v i d e d t h a t t h e i n p u t p o i n t s x ~ n d y J a r e r a t i o n a l , p a n d a a r e a l so r a t i o n a lw i t h l e n g t h s p o l y n o m i a l l y b o u n d e d i n t h e l e n g t h o f t h e i n p u t . F o r g i v e n ( p, a ),

c h e c k i n g t h a t t h e n u m b e r o f m i s c l a s s if i c a ti o n s d o e s n o t e x c e e d K c a n b e p e r -

f o r m e d i n t i m e p r o p o r t i o n a l t o t h e l e n g t h o f (p , a ), th e l e n g t h o f t h e i n p u t a n d

t h e n u m b e r r + s o f p o i n t s i n X u Y, i.e . i n p o l y n o m i a l ti m e w i t h r e s p e c t t o th e

i n p u t s iz e, t h e n u m b e r o f b i ts r e q u i r e d t o e n c o d e t h e d a t a x i, i ~ I a n d yJ, j ~ J .

( b) D P 2 i s N P - c o m p l e t e .

T h i s r e s u l t w il l b e o b t a i n e d b y p o l y n o m i a l ( a c t u a l ly li n e a r ) r e d u c t i o n o f t h e c l o s e d

d e n s es t h em i s p h e r e p r o b l e m ( C D H ) , w h o s e N P - c o m p l e t e n e s s h a s b e e n p r o v e d

b y J o h n s o n a n d P r e p a r a t a [ 1 3 ] , t o D P 2 . T h e c l o s ed d e n s e s t h e m i s p h e r e p r o b l e m ,i n it s sa t is f ia b i li ty f o r m , c a n b e s t a t e d a s ( se e G a r e y a n d J o h n s o n [ 8 ] ) :

G i v e n a s e t W o f m p o i n t s i n R , d o e s t h e r e e x i st a v e c t o r q i n R a n d a

s c a l a r b s u c h t h a t a t l e a s t L p o i n t s i n W l ie o n t h e s a m e s id e o f ( o r o n )

t h e h y p e r p l a n e H = {q ' z = 0}, i .e .: ca rd {w ~ W l q 'w > 0} > L?

Page 19: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 19/29

No vel Approaches to the Discrimination Problem 535

O

~ 0

Fig 4: Solving CD H

C o r r e s p o n d i n g t o C D H w e f o r m u l a t e a n i ns ta n c e o f D P 2 w i th X = W w T a n d

Y = T w h e re T co n t a i n s m - L + 1 co p i e s o f t h e o r i g in , an d we s e t K = m - L .

T h e n a n y s o l u t i o n t o t h i s D P 2 m u s t b e a h y p e r p l a n e g o i n g t h r o u g h t h e o r i g i n

a = 0 ) an d l eav e a t m o s t m - L p o in t s o f X m i s c l a s si f ied , i.e . t h a t a t l e a s t L p o i n t s

i n X e q u i v a l e n t l y L p o i n t s in W ) li e o n t h e s a m e s i d e o f H , t h u s s o l v i n g C D H

see f igure 4 ). [ ]

3 .3 S o l u t i o n A l g o r i t h m s

M i n i m i z i n g t h e n u m b e r o f m i s c la s s if i ed p o i n t s c a n b e a c h i e v e d b y s o l v in g t h e

b i le v e l p r o g r a m s 1 2). H o w e v e r , f o r n = 2 , a n O r + s ) 2 l g r + s ) ) a l g o r i t h m i s

a v a il a b le , b a s e d o n t h e f a ct t h a t t h e d i s c r im i n a t i n g h y p e r p l a n e h a s t o g o t h r o u g h

a t l e a s t t w o p o i n t s o f X w Y. I n i t s b as i c f o r m t h i s a l g o r i t h m c a n b e e x p r e s s e d a s

A l g o r i th m 2 - P o l

fo r a ll p o in t s z i n X w Y o

Page 20: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 20/29

536 P. M arcotte and G. Savard

0. S e t S = X u Y - { z } .

1. S e l e c t a n a r b i t r a r y a x i s g o i n g t h r o u g h t h e p o l e z.

2 . S o r t a l l p o i n t s i n S a c c o r d i n g t o t h e i r p o l a r a n g l e w i t h r e s p e c t t o t h e a x i s .

3 . S e l e c t a n a r b i t r a r y p o i n t z I i n S a n l e t H b e t h e h y p e r p l a n e s t r a i g h t l in e )

g o i n g t h r o u g h z a n d z 1.

4 . L e t N d e n o t e t h e n u m b e r o f m i s c l a s s i fi c a ti o n s w i t h r e s p e c t t o H .

5. le t S = S - {z 1 }.

i f S = ~ t h e n s t o p a n d o u t p u t N

ls s e t z I t o t h e p o i n t i n S c l o s e s t t o H a n d r e t u r n t o 4 .

I n a lg o r i th m 2 - P O L , s t e p 2 h a s c o m p l e x i t y O r + s) l g r + s )) . S ince , a f t e r s t e p

5 h a s b e e n p e r f o r m e d , t h e n u m b e r o f m i s c l a s s if i c a ti o n s c a n o n l y i n c r e a s e o r

d e c r e a s e b y 1 o r 2 , s t e p 4 c a n b e i m p l e m e n t e d i n c o n s t a n t t im e , a n d t h e o v e r a l l

w o r s t -c a s e c o m p l e x it y o f 2 - P O L is O r + s ) 2 l g r + s )) , a s p r e v i o u s l y c l a i m e d .

R e m a r k s : 1. I n t h e r e v e r se c o n v e x c a s e , 2 - P O L is a c o m p l e t e e n u m e r a t i o n s c h e m e .

I t is p o s s i b l e to i n c r e a s e i ts e ff ic i e nc y b y u s i n g s y m m e t r y r e l a t io n s h i p s a n d t h e

f a c t t h a t w e c a n r e st r ic t o u r s e a r c h t o h y p e r p l a n e s g o i n g t h r o u g h c o u p l e s x , y )

w i t h x ~ X a n d y ~ Y . A l s o , if p o i n t s u s e d a s p o l e s a r e s e l e c t e d in a s u i t a b l e o r d e r

c l o se n e i g h b o r s ) t h e n t h e s o r t in g s t e p 2 c a n b e i m p l e m e n t e d u s i n g r e o p t i m i z a t i o n

t e c h n i q u e s .

2 . A l g o r i t h m 2 - P O L c a n b e g e n e r al iz e d t o p r o b l e m s i n R n t o y ie ld a n

O r + s ) ~ l g r + s )) a l g o r i t h m , f o l l o w i n g t h e li n es d e v e l o p e d p r e v i o u s l y a n d t h e

r e c u rs iv e p r o c e d u r e m e n t i o n e d i n t h e p a p e r b y J o h n s o n a n d P r e p a r a t a 1 1 3] f o r

t h e d e n s e s t h e m i s p h e r e p r o b l e m . H o w e v e r s u c h a lg o r i th m s a r e n o t c o m p e t i ti v e

w i t h i m p l i c it e n u m e r a t i o n s c h e m e s i n h i g h d i m e n s i o n s , a s w e w i ll s ee .

3 .4 N u m e r i c a l R e s u l t s

W e c o m p a r e d t he ef fic ie nc y o f t h e b ile v el a p p r o a c h a n d t he 2 - P O L a n d N - P O L

e n u m e r a t i o n s c h e m e s fo r s o lv i n g D P 2 o n t h e se ts o f r a n d o m l y g e n e r a t e d p r o b -

l e m s p r e v i o u s l y d e s c r i b e d i n s e c t i o n 2 . 4 . A g a i n , n o a t t e m p t s h a v e b e e n m a d e t o

s p e c ia l iz e t h e B L P c o d e t o e x p l o it t h e s t r u c t u r e o f t h e b il e v e l p r o b l e m 1 2).

I n t a b l e s 5 , 6 a n d 7 w e p r e s e n t t h e n u m e r i c a l r e su l ts . N o t e t h a t , s in c e a l g o r i th m s

2 - P O L a n d N - P O L a re b a s e d o n e x p l i c i t i n c o n t r a s t w i t h implic i t ) e n u m e r a t i o n ,

s o l u t i o n t i m e s f o r t h o s e a l g o r i t h m s s h o u l d t h e o r e t i c a l l y b e e q u i v a l e n t f o r p r o b -

l e m s w i t h i d e n t i c a l r , s a n d n v a l u e s , a l b e i t d i ff e r e n t X a n d Y d a t a . T h e r e f o r e , f o r

g i v e n s a m p l e s iz e s, t h o s e 2 a l g o r i th m s w e r e n o t t e s t e d o n a l l t e n r a n d o m l y g e n e r -

a t e d s u b p r o b l e m s , a s i n g le o n e b e i n g s u ff ic ie n t t o s h o w t h e b e h a v i o r o f t h e m e t h o d s .

Page 21: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 21/29

N o v e l A p p r o a c h e s t o t h e D i s c r i m i n a ti o n P r o b l e m 5 7

Co n seq u en t ly , m ed ian an d s t an d a r d d ev ia t io n f ig u r es a r e n o t g iv en , b e in g eq u a l

to th e av e r ag e an d th e n u m b er ze r o r e sp ec tiv e ly .

I n v i ew o f t ab l e 5 , i t i s c l ea r t h a t 2 - P O L o u tp e r f o r m s th e b i l eve l ap p r o ac h f o r

so lv in g d i sc r im in a t io n p r o b le m s o n th e p l an e (n = 2). I n d eed , w h a tev e r t h e

d eg r ee o f o v e r l ap b e tw een X an d Y, p r o b lem s in v o lv in g h u n d r ed s o f p o in t s i n

each g roup cou ld be so lved wi thou t d i f f icu l ty . Never the less th is approach fa i l s

in d im en s io n s h ig h e r t h an 3 . I n d im en s io n 4 i t i s o u tp e r f o r m ed b y th e b i l ev e l

a lg o r i th m o n a ll t e st p r o b lem s , w i th th e ex cep t io n o f t h o se p r o b lem s co r r e -

spo ndin g to the f i r s t se t o f d ispers ion coef ficien ts . In d im ensions h igher th an 4 ,

a l g o ri th m N - P O L b e c o m e s i n tr a ct a b le . A r o u g h c a l c u l a ti o n s h o w s t h a t s o m e

2 x l0 s seco n d s w o u ld b e r eq u i r ed f o r so lv in g a D P 2 o f size 4 0 x 4 0 an d

d im en s io n 5 . Fo r so lv in g a D P 2 in v o lv in g th e sam e n u m b er o f p o in t s , 2 x 1 0 x3

s e c o n d s o f C P U t im e w o u l d b e r e q u ir e d i n d i m e n s i o n 1 0.

I n th e b i lev e l ap p r o ach , t h e p r o b lem s d i ff icu l ty i s d i r ec t ly p r o p o r t io n a l t o t h e

n u m b er o f n o d es ex p lo r ed in t h e b r an ch - an d - b o u n d t ree an d , a s i n d ica t ed in

sec t io n 2 .4 , i t i s r e l a t ed to t h e d eg r ee o f o v e r l ap b e twe en th e two g r o u p s o f p o in ts .

F o r p r o b lem s in v o lv in g a l a rg e n u m b er o f m i sc la s s if ied p o in t s , t h e av e r ag e

n u m b e r o f b r a n c h - a n d - b o u n d n o d e s c a n v a r y b y a f a c to r la r ge r t h a n 1 00 0 f o r

p r o b lem s o f s im i la r d im en s io n s g en e r a t ed th r o u g h d i ff e ren t m ea n s an d v a r i an ce-

cov ar ianc e m at r ices . F or the se t o f p rob lem s hav ing d ime nsions (40 , 20 , 4 ) ( see

tab le 7 ) th is fac to r go t as la rge as 1157 , whi le i t was a lways lower than 23 fo r

D P1 . Th i s can b e ex p la in ed b y th e f ac t th a t t h e n u m b er o f m i sc la s s if ica t io n s is

a f fec ted , t o a g r ea t e r ex t en t t h a t t h e o p t im a l v a lu e o f DP1 , b y th e d eg r ee o f

o v e r l ap b e tw een X an d Y. I n co n t r a s t t h e o b jec t iv e D P 2 is v e r y m u ch d ep en d en t

o n th e d eg r ee o f d i sp e r s io n o f t h e sam p le p o in t s.

These p re l imina ry resu l t s a re encourag ing , espec ia l ly in v iew of those resu l t s

o n DP 2 o b ta in ed o n m a in f r am e co m p u te r s f o r p r o b lem s o f s ize r = s = 5 0, n = 3

a n d r e p o r t e d i n t h e p a p e r s b y K o e h l e r a n d E r e n g u c [ 1 7 ] a n d S t a r e a n d

Jo ach im s th a le r [ 2 2 ] .

Conc lus ion

I n th i s p a p e r we in v es tig a ted , t h eo r e t i ca lly an d n u m er i ca lly , tw o f o r m s o f a

n o n p a r am e t r i c d i sc r im in a t io n p r o b lem , f o r wh ich n ew, a l t e r n a tiv e f o r m u la t io n s

wer e g iv en , an d sev e r a l a lg o r i th m s p r o p o sed . P r e l im in a r y r e su l t s o n sm a l l t o

m ed iu m - s i ze p r o b lem s a r e en co u r ag in g an d we in ten d , a s a seq u e l t o t h is w o r k ,

to t a i lo r so m e o f t h e m o r e p r o m is in g a lg o r i th m s to th e p a r t i cu la r s t r u c tu r e o f

th e d i sc r im in a t io n p r o b lem . I n p a r t i cu l a r we b e l i eve th a t t h e in c lu s io n o f n ew

o p t im a l i ty te s t s an d b o u n d s in t h e b il ev el p r o g r am m in g p r o ce d u r e th a t ex p lo i t

t h e s t r u c tu r e o f D P 2 co u ld s ign i fi can t ly d ec r ease co m p u t in g t im es .

Page 22: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 22/29

Table D

P2

d

a

s

oo 

r

=

2

s= 

1

r

=

2

s= 

1

r

=

2

s

=

2

r

=

3

s=1

~x 

:Cx

~r n

0

0

0

5I I 2

0

01

0 I I 2

0

0

2

0I 4 2

0

0

4

0I 4 2

0

0

8

0I

12

0

0

1

0I

12

n

c

n

c

n

c

n

c

n

c

n

c

1

0

4

7

4

6

2

0

3

4

1

4

1

2

6

6

1

8

9

6

8

8

4

9

BLP 

sm 

4

3

2

2

2

8

1

6

2

0

9

1

1

2

6

7

8

7

4

6

1

4

6

9

1

4

9

4

2

7

2

1

7

9

4

0

1

9

0

3

1

7

2

PO

-

0

0

0

0

-

0

0

-

0

0

-

0

0

-

0

0

1

8

8

1

6

2

3

4

5

0

3

1

9

4

6

1

2

0

1

5

6

0

4

2

BLP 

Sm 

1

3

5

0

4

6

2

4

3

4

2

7

5

6

3

8

1

7

1

8

4

8

3

5

1

6

9

2

1

5

4

2

7

9

5

4

2

1

5

3

1

7

2

PO

rn

-

0

0

0

0

-

0

0

-

0

0

-

0

0

-

0

0

2

2

1

3

9

8

6

2

8

8

6

3

1

4

9

8

4

4

3

2

9

6

7

8

BLP 

sm 

1

5

9

3

6

9

4

8

4

0

2

2

6

8

51

2

1

1

7

4

1

3

4

1

1

8

7

4

4

8

6

3

9

7

5

3

3

6

7

6

2

2

PO

-

0

0

0

0

-

0

0

-

0

0

-

0

0

-

0

0

2

6

1

7

7

8

7

1

5

8

5

0

1

6

1

8

2

2

3

8

7

8

7

6

BLP 

s

9

8

8

5

6

7

5

5

5

4

5

7

7

0

6

7

2

0

3

8

6

2

5

5

1

1

0

4

3

0

3

3

2

9

9

7

2

1

3

5

4

3

2

PO

-

0

0

0

0

-

0

0

-

0

0

-

0

0

-

0

0

 < 

Page 23: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 23/29

r=3

BLP 

s~

s=2

me

4

9

53

15

14

4

43

12

11

80

82

1

13

18

14

71

75

1

12

32

33

27

38

2

29

66

84

46

52

5

79

16

18

16

14

1

18

2POL 

-

00

00

-

00

-

00

-

00

-

00

18

29

91

12

1

16

r=3

BLP 

s~

s=3

46

62

31

50

2

33

28

2

0

10

14

1

25

10

17

61

15

8

18

6

0

8

5

14

2

2

5

78

24

43

31

47

1

26

2POL 

-

00

00

-

00

-

00

-

00

-

00

26

33

28

31

9

19

4

6

61

15

2

6

4

67

16

12

76

13

1

13

42

60

30

59

2

40

12

2

3

92

19

1

1

 

r=4

BLP 

s

s=2

me

12

23

27

47

5

87

2POL 

-

00

00

-

00

-

00

-

00

-

00

2

2

47

14

29

2

41

58

17

35

66

4

90

10

27

57

12

9

2

6

r=4

BLP 

s~

s=3

9

6

1

0

25

4

3

8

1

9

38

6

4

14

11

2

54

16

22

72

14

9

16

2POL 

00

00

-

00

-

00

-

00

-

00

4

2

1

8

21

61

3

8

401

1

1

2

 

32

73

1

2

1

56

17

42

12

4

12

01

 

0.14 

16

4

7

10

29

1

38

01

r=4

BLP 

sm 

s=4

4

8

1

1

11

41

4

9

901

2POL 

12

48

10

37

1

33

01

Z O Q O ~

Page 24: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 24/29

Ta

e

6

D

P2

s

d

a

s

r 2

s= 

1

r

=

2

s= 

1

r

=

2

s

=

2

r

=

3

s=1

  g gn

0

0

0

5I I 3

0

01

0I I 3

0

0

2

0

I 4 3

0

0

4

0I 4 3

0

0

8

0I

13

0

0

1

0I

13

n

c

n

c

n

c

n

c

n

c

n

c

1

0

6

9

3

8

1

4

2

4

1

9

8

4

4

9

1

0

6

7

8

2

4

8

BLP 

sm 

1

8

4

7

6

1

2

7

4

2

4

1

8

5

4

1

7

7

0

1

2

5

3

m

e

8

4

6

1

6

7

7

3

8

3

2

4

3

2

1

3

2

1

N

-PO

-

5

2

5

2

-

5

2

-

5

2

5

2

-

5

2

4

2

2

5

1

8

7

5

1

8

6

6

2

2

1

3

5

6

3

4

1

6

7

8

BLP 

sm 

2

3

1

5

1

2

1

3

1

8

6

5

5

4

3

9

5

7

3

5

1

6

1

3

2

1

6

5

3

5

4

2

6

7

6

6

2

2

9

3

2

6

N

-PO

-

8

8

8

8

8

8

-

8

8

8

8

-

8

8

6

4

4

8

1

8

8

5

1

4

1

5

2

6

1

5

7

8

6

7

1

0

1

4

BLP 

sm 

4

5

3

3

8

8

6

1

1

4

9

5

3

7

2

6

6

4

4

1

1

6

1

4

3

2

0

8

6

8

1

1

6

1

1

6

6

5

9

3

4

1

N

-PO

-

1

9

1

9

-

1

9

-

1

9

1

9

-

1

9

5

4

5

8

1

4

1

6

8

6

8

4

5

4

7

1

2

4

2

4

3

2

4

2

BLP 

sm 

4

1

4

7

9

0

9

0

8

4

8

7

4

1

4

7

2

1

2

9

1

8

2

0

m

e

4

3

8

6

6

4

4

4

4

3

4

7

1

1

5

3

3

8

N

-PO

-

2

2

-

2

2

-

2

2

-

2

2

-

2

2

-

2

2

O  g

Page 25: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 25/29

r

=

3

BLP 

Sm 

S

=

2

me

N

-PO

rn

r

=

3

BLP 

sm 

s

=

3

me

N

-PO

r

=

4

BLP 

sm 

s

=

2

me

N

-PO

1

9

1

7

1

1

1

2

1

1

1

2

4

3

5

2

6

3

2

1

1

0

-

3

9

2

2

3

4

1

1

2

5

1

2

4

-

3

9

6

2

1

6

2

7

4

2

7

1

1

2

6

2

6

1

5

2

8

1

2

8

3

9

9

8

1

4

6

2

9

2

6

1

2

1

2

1

2

3

r

=

4

BLP 

sm 

s

=

3

me

9

2

9

3

5

1

3

1

0

7

7

3

9

1

2

1

1

9

0

7

2

1

8

7

3

9

3

3

5

0

-

3

9

4

4

7

3

2

7

4

4

3

5

9

6 0 

5 8 

4

8

5

2

1

9

9

8

6

4

7

3

9

7

3

9

0

3

4

-

6

7

-

6

7

6

7

6

7

-

6

7

-

6

7

1

0

1

4

3

0

5

1

3

2

5

4

4

8

8

9

1

2

1

1

2

0

3

9

6

5

9

7

2

8

3

1

3

8

5

4

6

5

9

9

9

6

1

1

3

6

6

0

8

1

3

2

4

8

2

3

3

2

3

0

7

1

2

9

1

3

7

9

-

7

9

-

7

9

-

7

9

-

7

9

7

9

7

2

1

2

1

4

2

2

2

5

1

7

2

1

1

4

0

8

7

5

1

0

4

4

8

2

1

0

2

3

4

8

4

8

1

1

0

4

3

9

1

6

1

9

N

-PO

9

8

-

9

8

-

9

8

9

8

1

4

2

3

1

7

2

6

5

1

9

2

0

6

9

1

3

4

0

1

4

5

1

2

3

4

1

4

3

0

5

1

9

r

=

4

BLP 

Sm 

S

=

4

me

9

0

2

8

5

0

1

2

7

1

1

 

9

8

9

8

4

8

1

3

1

8

5

6

3

4

9

2

2

0

7

4

3

9

4

9

2

8

-

1

6

-

1

6

_

9

8

2

4

6

4

1

9

6

1

5

-

1

6

1

6

1

6

N

-PO

-

1

6

ZQ,  O 

O :1

O c

Page 26: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 26/29

5 4 2 P M a r c o t t c a n d G S a v a rd

t ~

0

0

0

0

- t

0

~ i c 5 e , i

C ~ C b ~ ,

~ ~

9

~ . o . ~

~,.,i t.~l -xl

, 1 , .. .1 , . 1 , - 1

C b

I I It...

r 0

I I I II I I I I I I

Page 27: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 27/29

r=

3

BL

s

s=

2

2

2

3

5

1

3

1

6

3

4

0

1

8

2

7

2

0

3

6

5

7

3

2

2

3

5

3

8

4

4

9

1

5

62

1

5

75

1

3

3

52

7

2

1

8

1

1

2

9

1

2

4

30

56

23

36

30

46

N-POL

-

6

9

-

6

9

-

6

9

-

6

9

'

6

9

-

6

9

8

9

1

3

5

7

9

4

6

1

8

r=

3

BL

s~

s=

3

1

0

3

3

2

3

4

8

9

2

6

6

2

1

8

4

8

7

1

4

8

3

2

6

4

4

2

1

4

1

2

4

4

6

6

1

9

2

4

4

1

6

9

4

72

1

8

1

5

2

4

3

78

N-POL

-

1

4

-

1

4

1

4

-

1

4

1

4

-

1

4

40

1

7

39

67

3

8

6

9

1

4

2

0

3

8

5

4

2

5

4

4

6

0

1

7

4

7

8

7

3

4

6

1

2

2

4

5

r=

4

BL

sm 

32

81

29

58

s=

2

N-POL

r=

4

BL

sm 

s=

3

N-POL

r=

4

BL

sm 

s=

4

N-POL

3

5

2

3

6

1

1

3

1

3

66

6

1

8

3

59

-

1

7

-

1

7

1

7

-

1

7

-

1

7

-

1

7

 

-

4

4

1

2

7

0

1

0

58

1

7

1

6

4

1

38

1

0

-

-

5

0

1

1

6

3

1

6

~

1

0

2

2

4

3

25

74

 

-

2

5

1

4

1

0

3

1

9

1

3

1

3

73

 

2

4

-

2

4

2

4

-

2

4

-

2

4

-

 

2

4

m  m 

1

8

,3

9

1

7

~3

4

8

2

2

I

8

8

5

8

9

4

2

2

1

5

6

4

0

1

7

8

3

2

1

9

3

6

5

2

1

4

4

4

1

8

2

8

8

1

0

3

7

2

5

~

6

4

3

~

92

-

5

2

-

5

2

5

2

-

5

2

-

5

2

-

5

2

ZO > O g O to

Page 28: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 28/29

544 P. Marcotte and G. Savard

Fina lly i t would be interest ing to co mpar e the practical performance of the

proposed bi level prog ramm ing approac h to that of more es tabl ished methods of

global optimizat ion, such as the cutt in g-plan e meth ods of Horst an d T uy [12].

Acknowledgments: We would like to thank Grrald Marquis and Xavier Haurie for their help in

conducting the numerical experiments.

eferences

[1] Bajgier SM and Hill AV 1982) An experimental comparison of statistical and linear program-

ming approaches to the discriminant problem. Decision Sciences 13:604-618

[2] Bodlaender HL, Gritzmann P, Klee V and Van Leeuwen J 1990) Computational complexity

of norm-maximization.Combinatorica 10:203-225

[3] Cavalier TM, Ignizio JP and Soyster AL 1989) Discriminant analysis via mathematical pro-

gramming: certain problems and their causes.Com puters and Operations Research 16:353-362

[4] Fisher R A 19 36) Th e use of mu l t ip le measure men ts in taxono mic problems. Annals o f Eugenics

7:179-188

[5] Freed N and Glover F 1981) A linear programming approach to the discriminant problem.Decision Sciences 12:68-74

[6] Freed N and Glover F 1981) Simple but powerful goal programming models for discriminant

problems. European Journal of Operational Research 7:44-60

[7] Freund RM and Orlin JB 1985) On the complexity of four polyhedral set containment

problems. Mathematical Programming 33:139-145

I-8] Garey MR and Johnson DS 1979)Computers and intactahility, Freeman WH, San Francisco

[9] Gaither N and Glorfeld W 1982) On using linear programming in discriminant problems.

Decision Sciences 13:167-171

[10] Glover F 1990) Improved linear programming models for discriminant analysis. Decision

Sciences 21 : 771-785

[11] Hansen P, Jaumard B and Savard G 1990) New branch-and-bound rules for linear bilevel

programming, Cahier du GERAD G-89-09. Ecole Polytechnique, Montrral, submitted forpublication

[12] Horst R and Tuy H 1990) Global optimization. Deterministic approaches. Springer-Verlag,

New-York

[13] Johnson DS and Preparata FP 1978) The densest hemisphere problem. Theoretical Computer

Science 6: 93 - 107

[14] Kojima M, Mizuno S and Noma T. A polynomial-time algorithm for a class of linear comple-mentarity problems, to appear in Mathematical Programming

[15] Khachyan LG 1979) A polynomial algorithm in linear programming in Russian),Doklady

Akademi ia Nauk SSSR 244:1093-1096; English translation: Soviet Mathematics Doklady 20:191-194

[16] Koehler GJ 1989) Characterization of unacceptable solutions in LP discriminant analysis.Decision Sciences 20: 239-257

[17] Koehler GJ and Erenguc SS 1990) Minimizing misclassifications in linear discriminantanalysis.Decision Sciences 21:63-85

[18] Konno H 1976) Maximization of a convex quadratic function subject to linear constraints.Mathematical Programming 11 : 117-127

Page 29: Novel Approaches Discrimination

8/12/2019 Novel Approaches Discrimination

http://slidepdf.com/reader/full/novel-approaches-discrimination 29/29

No v e l Ap p r o a c h es t o t h e D i s c r im in a t i o n P r o b l e m 5 45

[19] L in YY a nd Pan g JS (1987) I te ra t ive me thods fo r large convex quadra t ic p rograms: a su rvey .

S I M Journal on Control and Optimization 2 5 :3 8 3 - 4 1 1

[20] M arkow sk i CA an d M arkow sk i EP (1987) An exper imen ta l com par ison o f severa l approaches

to t h e d i s c r im in a n tp rob le m w i th bo th qua l i ta t ive and quan t i ta t ive var iab les . European Journalo f O perational Research 2 8 : 7 4 - 7 8

[ 2 1 ] Ru b in P A ( 19 8 9) E v a lu a t in g t he m a x im iz e m in im u m d i s ta n c e f o r m u la t i o n o f t h e l i n e a r

d i s c r im in a n t p r o b le m . European Journal o f Operational Research 4 1 :2 4 0 - 2 4 8

[22] S tam A a nd Joach im stha le r EA (1990) A com par ison o f a rob us t mixed- in teger approach to

exist ing m etho ds for e stablishing classif ication rules for the disc r im inan t problem . European

Journal o f O perational Research 4 6 : 1 1 3 - 2 2

[23] Wolfe P (1976) Fin din g he nearest point in a polytope. Mathem atical Programming 11 : 128-14 9

Received: Fe bru ary 1991

Revised ve rsion received: Octo ber 1991