Data Attribute Reduction using Binary Conversion

Transcription

1 Fengmng M. Chng Dt Attrute Reducton usng Bnry Converson FENGMING M. CHANG Deprtment of Informton Scence nd Applctons As Unversty Wufeng, Tchung 4354 Twn Astrct: - Whle lernng wth dt hvng lrge numer of ttrute, system s esy to freeze or shut down or run for long tme. Therefore, the proposed Bnry Converson (BC) s novel method to solve ths knd of lrge ttrute prolem n mchne lernng. The purpose of BC s to reduce dt dmensons y nry converson process. All the ttrutes re reserved ut comned nto few numers of new ttrutes nsted of tht some ttrutes re removed. To prevent the nformton loss prolem durng the converson, ech nry type dt vlue occupes ts own dgtl poston n BC. In ddton, 4 dt sets: nuses, ACLP, MONK3, nd Buseskod dt re used n ths study to test nd compre the lernng ccurces nd lernng tme. The results ndcte tht the proposed BC cn keep out the sme level of ccurcy ut ncrese the lernng effcency. Key-Words: - Bnry converson, Lrge ttrute, Mchne lernng, Neuro-fuzzy, Meg-fuzzfcton Introducton Recent hve shown wdely pplctons n the felds of Artfcl ntellgence (AI) for dt clssfcton or predcton [-4]. Mny methods were proposed. For most of the exmples n prevous studes, the numer of nput ttrutes s not lrge. They proly only provde theoretc model for reserchers. However, rel dt n some theoretc studes nd some prctcl pplctons hve plenty of nput ttrutes. It cuses some prolems. Frst, some systems wll esly shut down ecuse the clcultons of the mchne lernng re too lrge. Second, some lernng progrms hve ther lmts. The lernng methods tht mostly need to reduce nput ttrute numers re Artfcl Neurl Network (ANN), Fuzzy Neurl Network (FNN, neuro-fuzzy), nd Meg-fuzzfcton, the lter s mproved sed on FNN. FNN nd Megfuzzfcton re more dffcult to perform thn ANN. FNN dels wth the network lernng usng fuzzy memershp functons. Becuse the defuzzfcton clcultons n FNN re complex nd dffcult, most of the tme fuzzy memershp functons re setup s trngulr, generlzed ell, trpezodl, nd so on so tht they re esy to clculte. Anomlous shpes fuzzy memershp s not recommended ecuse t s lmost mpossle to defuzzfy usng progrms eforehnd even though the defuzzy clculton s stll not effcent. When the nput ttrute mounts re lrger thn 6, the FNN progrm could not perform normlly. Most of the tme, the computer went on hold wthout ny response, t froze. In ths rtcle, nuses dt set hs 9 nput ttrutes. These dt re used s fl for lernng when usng FNN or Meg-fuzzfcton methods ecuse they hve too mny ttrutes. On the other hnd, some mchne lernng progrms hve upper lmt of network nodes, nd reducton of the nput ttrute mounts s lso necessry. Ltertures Revew The reltve works nd mchne lenng methods tht re used for comprson re revewed n ths secton.. Dt ttrute reducton It s wy to mprove mchne lenng effcency y reducng numer of dt ttrutes. In the reltve reserches, Shen nd Chouchouls proposed the frst dt ttrute reducton method, Rough Set Attrute Reducton (RSAR) method, to remove redundnt nput ttrutes for dscrete vlues from complex systems. Besdes, n pproxmte reducts concept ws offered y Beynon. In the men tme, he lso proposed Vrle Precson Rough Sets (VPRS) model to fnd out the smllest set of ttrutes [6]. Afterwrds, VPRS ws ISSN: Issue 7, Volume 8, July 009

2 Fengmng M. Chng ppled n other studes, such s Hsu et l. used VPRS model for mole phone test procedure [7], nd Inrn et l. used VPRS for feture selecton of we usge mnng [8]. In ddton, lthough Ang nd Quek dd not reduce dt ttruute ut reduce fuzzy rules y comned rough set nd neuro-fuzzy lernng to mprove mchne lernng effcency[9].. Byesn networks [3] Byesn networks (BN) s herrchcl clssfcton method. They re conssts of grph nd prolty theory s grphcl models []. BN s defned s drected cyclc grph or prolstc grphcl model tht presents set of vrles nd ther cusl nfluences. The cusl dependences etween the vrles re expressed y condtonl prolty dstrutons. It usully set up numerc vlues wth norml or Gussn dstruton n BN [3]..3 Decson tree ID3 nd C4.5 [3] ID3 s nother herrchcl clssfcton method to e decson tree [4]. A decson tree performs ctegorcl splt followng the vlue numer of nput ttrutes. A decson tree s predctve model mppng from oservtons to predct the trget vlues. Wrtten y C, C4.5 s n mproved verson of ID3. Both symolc nd numerc vlues of nput ttrutes cn e clssfed nd then outputs clssfcton tree..4 Support vector mchne [3] Bsclly SVM seprtes dt ponts nto two groups of lner seprle dt sets. It tres to fnd out the optml hyperplne y mnmzng n upper ound of the errors nd mxmzng the dstnce, mrgn, etween the seprted hyperplne nd dt [5]. It ws frst nvented for nry clssfcton prolems sed on sttstcl theory. A mxml seprtng hyperplne s ult y SVM to mp nput vectors to hgher dmensonl spce. Two prllel hyperplnes re ult nd the dt re seprted on ech sde of the hyperplne. Gven trnng dtset ( ) y x,, = K,, k, where {, } k n x R nd y, SVM tres to fnd out the optml soluton prolem usng the followng form mnly: mn w,,ε T w w + C k = suject to T y ( w φ( x ) + ) ε, ε 0 ε.5 Artfcl neurl network [3] An ANN conssts of nodes tht re nterconnected usng mthemtcl model wth computtonl model. It computes nputs dt y ove models to gn ts output. In more prctcl terms, n ANN s non-lner tool. It cn model complex reltonshps etween nputs nd output dt. In most of tmes, n ANN s n dptve system tht cn djust the prmeters to mprove ts performnce for gven tsk. There re three types of neurl network lernng lgorthms: supervsed, renforcement, nd unsupervsed lenng. Bckpropgton neurl network (BPN) s the est known supervsed lernng lgorthm..6 Neuro-fuzzy [6] The ANFIS [0], fuzzy neurl network (FNN) tool, performs neurl lernng usng fuzzy typed numers. Gven set of nput nd output dt, the ANFIS cn constructs fuzzy nference system wth memershp functons vlues dpted usng ckpropgton lgorthm or n comnton wth lest squre method. The dptton functon of the ANFIS provdes the mchne lernng system wth FNN chrcters. The sc model of the ANFIS s the Sugeno fuzzy model []-[4]. In the model, ssumng x nd y re two nput fuzzy sets nd z s the output fuzzy set, nd the fuzzy f-then rules s formtted s: If x = P nd y = Q then z = f ( x, y) Consder two frst-order rules of Sugeno fuzzy model, the f-then rules cn e: where m, P, n, Rule A : If Rule B : If x = P nd Q y = f = m x + n y + c, x = P nd Q y = f = m x + n y + c Q, P, nd c, m,, then, then Q re fuzzy set vlues nd c re constnts. In n, nd fgured presentton, Fg. 3 shows ANFIS structure wth fve-lyer rtfcl neurl network. Denote tht the output of the th node of lyer l re O,. In Lyer of the ANFIS, l O = ( ), =,, or, µ M x ISSN: Issue 7, Volume 8, July 009

3 Fengmng M. Chng where O µ M nd = µ ( ), =3,4, N y µ N re fuzzy memershp functons tht cn e ny memershp type such s trngulr or generlzed ell functon. x y A B A B X Lyer Y Lyer Lyer 3 Lyer 4 Fg. 3. The ANFIS structure. Lyer 5 For nodes n Lyer, the outputs w re the product of the outputs of lyer nd re used s the weghts of Lyer 3: O = w = µ ( x) ( ), =,, M µ N y In Lyer 3, the output of every node s normlzed y clculton s the followng: O 3, w = w =, =,. w + w Next, Lyer 4 s the defuzzy lyer whch dpts node vlues wth equton: O = w f = w ( m x + n y + c ), for =, 4, where m, n, nd c re consequent prmeters of the nodes. Fnlly, the ffth lyer s to compute the output of ll the nput sgnls usng the equton: w f 5 = w f =, for =, w O,.7 Meg-fuzzfczton [9, 7, 8, 9, 30, 3] f x f y Meg-fuzzfcton method ws proposed for the purpose of solvng the predcton of the est strtegy prolem n the Flexle Mnufcturng System (FMS) when dt re smll [, 4, 6]. In the studes of [, 4, 6] for the Meg-fuzzfcton, severl concepts were offered, such s dt fuzzfed, contnuous dt nd, domn externl expnson, nd dt s dptton. The concept of the contnuous dt nd ws frst proposed y L, et l. []. Such dt contnung technology ms to fll the gps etween ndvdul dt nd mke ncomplete dt nto the more complete sttus s presented y Hung nd Morg [-]. Furthermore, n the studes of L, et l. nd Chng [, 4, 6], the domn rnge externl expnson concept ws lso proposed nto the procedure of the contnuous dt nd method. In ddton to fllng the dt gps wthn the dt set, the purpose of domn externl expnson s lso to dd more dt outsde the current dt lmts, ecuse possle dt re expected not only nsde ut lso outsde the current dt rnge. To fll the gps etween crsp dt, crsp lernng dt re trnsformed nto contnuous dt s Fgure llustrtes. In Fg., there re fve orgnl crsp dt. When these dt re trnsformed nto contnuous type, vrtul dt etween the crsp dt re thus generted. crsp dt contnuous dt Fg.. Crsp dt re trnsformed nto contnuous The dt effects hve to e estmted lso. In ths reserch, contnuous dt re presented n fuzzy memershp functon forms. The fuzzy memershp functon cn e generl ell, trngle, or even nomlous type nd s the dt s effectve weght n the FNN lernng lter. Most of the tme, n symmetrc fuzzy memershp functon s ntted. For exmple, trngulr fuzzy memershp cn e: µ ~ A 0, x < mn x mn, mn x md md mn ( x ) = mx x, md x mx mx md 0, x > mx ISSN: Issue 7, Volume 8, July 009

4 Fengmng M. Chng where mn s the mnmum vlue nd mx s the mxmum vlue of the crsp lernng dt, nd mn+ mx md =. When crsp re trnsformed nto contnuous, oundres of the contnuous dt nd re mnmum nd mxmum vlue of the orgnl crsp dt. However, the rel dt rnge s possle outsde ths dt nd. In order to uld up rel knowledge, the dt nd s externlly expnded to the possle rnge n ths study s llustrted n Fg. 3. mnmum Orgnl lernng dt Fuzzy memershp functon () mxmum frst to the thrd ttrutes to e the frst new ttrute. On the left of the Fg. 4, 3 decml numers, 4, 3, nd, re trnsferred nto 3 Boolen numers ccordngly. Consderng the mxmum vlue of ech ttrute, Boolen numer for decml numer should e 0. So tht the numer of t n Boolen numer for ech ttrute s fxed. Next, the 3 Boolen numers re physclly comned to e unque Boolen numer s shown n the mddle prt of Fg. 4. Ech orgnl Boolen numer occupes ts own dgtl poston n the comned Boolen numer formt wthout mxng wth other numers. After tht, for the convenence of clculton n the rel world, ths comned Boolen numer s trnsferred to decml numer. In the ove process, the 3 nput vlues re comned nto unque decml vlue 77 nd the nput ttrutes re reduced. The reson for not comnng the decml numer drectly, such s comne 4, 3, to e 43 s ecuse tht 43 s gger thn 77, the result of BC. Smller numer s eser for clculton. 3 decml numers Trnsferred to 3 Boolen numers Comned 6 Boolen numers to one Trnsferred to decml numers new mnmum () new mxmum Fg. 3. Two trngulr memershp functon vlues: () efore, () fter domn rnge externl expnson Becuse dt hve een trnsformed nto contnuous dt nd, FNN re used n ths study. Durng the lernng process, the shpe of fuzzy memershp functon s dpted to ft the est lernng results. 3 The Proposed Method Frst, dt set of nuses [0] tht conssts of 9 nput nd output ttrutes s used to expln the proposed BC method frst. The nuses cn not e performed well n FNN nd meg-fuzzfcton methods. Vlues of ts 0 ttrutes re ntegers. The vlues of the frst ttrute re {,, 3, 4}, vlues of the second, the 8 th, nd output ttrutes re {,, 3}, nd vlues of the other ttrutes re {, }. The process of the BC method s smple. Ech decml numer cn e trnsferred nto Boolen numer one on one mppng. For exmple, n the frst nstnce of our nuses dt, 9 ttrutes re comned nto 3 new ttrutes. We comne the Fg. 4. The process of the Boolen Converson. Wth Fg. 4 s n exmple, there re 3 nputs were converted nto sngle new nput. Frst, the orgnl 3 nputs {4, 3, } re converted nto Boolen dgt numer tht s {00,, 0}. Second, these 3 Boolen dgt numers re physclly comned nto one Boolen dgt numer: 000. The correspondng decml numer s 77. It cn e expressed y the nry system s: { 00,, 0} 000 = 00* *00 + 0* It could e Boolen weght expressed y Boolen system s: B = [ ] ISSN: Issue 7, Volume 8, July 009

5 Fengmng M. Chng Tle. An explnton of convertng 9 ttrutes to 3 new decml ttrutes. # # #3 #4 #5 #6 #7 #8 #9 Orgnl decml vlue Converse to Boolen vlue Comne to three Boolen vlue Converse to three decml vlue or expressed y decml system: 4 0 { 4, 3,} 77 = 4 * + 3 * + * nd the nry weght vector s B = [ 4 4 Results nd Comprsons In ths study, 4 dt sets re offered to check lernng ccurces nd tme of the results of oth non-pplyng nd pplyng BC. These dt sets re nuses, ACLP, MONK3 nd Buseskod dt sets. 4. nuses dt The nuese dt tht hs een mentoned n secton 3 s used y the proposed method n ths susecton. Tle shows one record of the dt. There re 9 nput nd one output ttrutes n the dt. The st to the 3 rd ttrutes re converted nto new nput ttrute, the 4 th to the 6 th ttrutes re comned nto the nd new nput y BC, nd the 7 th to the 9 th re comned nto the 3 rd new one. Therefore there re only three new nput ttrutes. As shown n Tle, the new nput record s {77, 38, 9}. After ll ttrutes re converted usng BC, the dt re tested nd compred usng BN, C4.5, SVM, ANN, FNN, nd Meg-fuzzfcton methods wth 0-folds cross-vldton testng. Ech fold re used 0 ] s testng dt n turn nd the remnng totl of 9 folds dt re used s trnng dt. The results re presented n Tle. Wthout usng BC, FNN nd Meg-fuzzfcton fl to perform. After pplyng BC, t cn esly perform mchne lernng usng FNN nd Meg-fuzzfcton methods. Most of the predcton ccurces fter usng BC re even lttle hgher thn wthout usng t n ths cse, nd lernng tme decreses. Fg. 5 compres the predcton ccurces under dfferent lernng methods. Fg. 6. llustrtes the lernng tme. For nuses dt, even fter usng BC, tme for FNN nd meg-fuzzfcton s stll lrge. For other lernng methods, lernng tme reduced. 4. ACLP dt There re n totl 40 nstnces n ACLP dt wth 6 nput nd output ttrutes. The st, the 5 th, nd the 6 th ttrutes hve vlues of {,, 3}, nd the other ttrutes hve vlues of {, }. For neuro-fuzzy lernng, 6 nput ttrutes cuse lernng process very slowly. Fortuntely, the vlues of ech ttrute re not lrge. We cn compre the results of FNN nd meg-fuzzfcton methods. The results re presented n Tle 3, Fg. 7, nd Fg. 8. Accurces fter usng BC re lttle lower thn efore, ut tme s svng. Before pplyng BC n FNN nd meg-fuzzfcton, tme for lernng s very lrge. After usng BC, tme s sved lrgely. Non-BC BC Tle. The comprson of nuses dt. Method Byesn C4.5 SVM ANN FNN Meg-fuzzfcton Accurcy 93.4 % 93.4% 86.84% 85.53% Tme(sec) Fl to perform Fl to perform Accurcy 94.74% 93.4% 84.% 9.% 95% 95% Tme(sec) ISSN: Issue 7, Volume 8, July 009

6 Fengmng M. Chng ccurcy 00.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 0.00% 0.00% 0.00% Byesn C4.5 SVM ANN FNN megfuzzfcton ccurcy wthout usng BC ccurcy fter usng BC Fg. 5. The ccurcy comprson efore nd fter usng BC method y sx methods for nuses dt. 3.5 tme Byesn C4.5 SVM ANN FNN megfuzzfcton lernng tme wthout usng BC lernng tme fter usng BC Fg. 6. The lernng tme comprson efore nd fter usng BC method y sx methods for nuses dt. Tle 3. The comprson of ACLP dt. Method Byesn C4.5 SVM ANN FNN Meg-fuzzfcton Non-BC BC Accurcy % 87.86% 90.7% 84.9% 84.73% 84.73% Tme(sec) Accurcy 80.7% 89.9% 8.43% 8.43% 80.38% 8.3% Tme(sec) ISSN: Issue 7, Volume 8, July 009

7 Fengmng M. Chng ccurcy 00.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 0.00% 0.00% 0.00% Byesn C4.5 SVM ANN FNN megfuzzfcton ccurcy wthout usng BC ccurcy fter usng BC Fg. 7. The ccurcy comprson efore nd fter usng BC method y sx methods for ACLP dt tme lernng tme wthout usng BC lernng tme fter usng BC Fg. 8. The lernng tme comprson efore nd fter usng BC method y sx methods for ACLP dt. 4.3 Monk3 dt Monk3 dt were creted y Sestn Thrun (see UCI Mchne Lernng Repostory []) whch hs 43 nstnces, 6 nputs nd output ttrutes. Becuse the numer of ttrutes s not lrge n ths cse, we cn compre the lernng ccurces of FNN nd meg-fuzzfcton wth nd wthout usng BC gn. Tle 4 shows the results. In ths cse, FNN nd meg-fuzzfcton cn e performed ut wste lrge tme efore usng BC. All the ccurces fter usng BC re lttle lower thn efore. The lernng ccurces re lso compred n Fg. 9 nd lernng tme s compred n Fg. 0. Stll, lenng tme efore usng BC for FNN nd meg-fuzzfcton s very lrge, ut ecomes very smll fter pplyng BC. Non-BC BC Tle 4. The comprson of Monk3 dt. Method Byesn C4.5 SVM ANN FNN Byesn C4.5 SVM ANN FNN megfuzzfcton Megfuzzfcton Accurcy 9.36 % 00% 80.56% 00% 00% 00% Tme(sec) Accurcy 89.% 96.06% 76.90% 98.87% 97% 98% Tme(sec) ISSN: Issue 7, Volume 8, July 009

8 Fengmng M. Chng ccurcy 00.00% 90.00% 80.00% 70.00% 60.00% 50.00% 40.00% 30.00% 0.00% 0.00% 0.00% Byesn C4.5 SVM ANN FNN megfuzzfcton ccurcy wthout usng BC ccurcy fter usng BC Fg. 9. The ccurcy comprson efore nd fter usng BC method y sx methods for Monk3 dt tme Byesn C4.5 SVM ANN FNN megfuzzfcton lernng tme wthout usng BC lernng tme fter usng BC Fg. 0. The lernng tme comprson efore nd fter usng BC method y sx methods for Monk3 dt. 4.4 Buseskod dt There re 76 nstnces n Buseskod dt set wth 8 nputs nd one output ttrutes. The vlues of the st to the 7 th nputs re {0, } nd the vlue rnges of the 8 th nput re {0,, } wth the vlues of output {, }. Stll, the lernngs of FNN nd Megfuzzfcton fl to perform efore ppled BC. The comprson results re shown n Tle 5, Fg., nd Fg.. The lernng ccurces fter BC re lttle worse or equl thn those efore BC, ut the lernng effcences re mproved fter BC. Tle 5. The comprson of Buseskod dt. Method Byesn C4.5 SVM ANN FNN Meg-fuzzfcton Non-BC Accurcy 00% 98.68% 00% 00% Tme(sec) Fl to perform Fl to perform BC Accurcy 99.05% 98.68% 98.4% 00% 00% 00% Tme(sec) ISSN: Issue 7, Volume 8, July 009

9 Fengmng M. Chng ccurcy 00% 90% 80% 70% 60% 50% 40% 30% 0% 0% 0% Byesn C4.5 SVM ANN FNN megfuzzfcton ccurcy wthout usng BC ccurcy fter usng BC Fg. 0. The ccurcy comprson efore nd fter usng BC method y sx methods for Buseskod dt. 5 4 tme 3 0 Byesn C4.5 SVM ANN FNN megfuzzfcton lernng tme wthout usng BC lernng tme fter usng BC Fg.. The lernng tme comprson efore nd fter usng BC method y sx methods for Buseskod dt. 5 Conclusons In ths study, novel BC method s proposed to del wth the prolem of tht dt wth lrge numer of ttrutes my cuse system freezes or shuts down. BC reduces ttrutes numer y comnng some of the ttrutes nto smller numer of new ttrutes nsted of tht removng some ttrutes from dt. After ttrutes re comned nd reduced, lernng ccurces nd lernng tme re compred y BN, C4.5, SVM, ANN, neuro-fuzzy, nd Megfuzzfcton lernng methods. Dfferent lernng methods hve dfferent lernng chrcterstcs, nd good lernng method for one dt s possle not sutle for nother dt [3]. The purpose of the comprson for neuro-fuzzy, such s BN, C4.5, SVM, nd ANN, s to compre the lernng ccurces nd effcences y BC ecuse FNN nd Meg-fuzzfcton methods cn not e performed n lrge ttrute dt. The purpose of the proposed BC s not to mprove lernng ccurcy ut to solve the prolem of flng to perform n lrge ttrute dt. Hence, lernng ccurces re not mproved n some dt sets, ut BC ecomes method for lrge ttrute condton nd lernng ccurces re not too worse thn efore BC. In ths study, 4 dt sets, nuses, ACLP, MONK3, nd Buseskod re used to test nd compre the lernng results. Some of ther lernng ccurces fter usng BC re lttle lower thn efore, some hve lttle hgher ccurces. In generl, the lernng ccurcy fter pplyng BC s not worse. In ddton, lenng tme s shortened fter BC s used. Fcng the prolem of fl to perform n neuro-fuzzy, the proposed BC method ndeed solves the prolem of dt hve lrge ttrutes n lernng n ref. Acknowledgement ISSN: Issue 7, Volume 8, July 009

10 Fengmng M. Chng Thnks re due to the support n prt y the Ntonl Scence Councl of Twn under Grnt No. NSC H MY. References: [] Y.Y. Yo, Grnulr computng: sc ssues nd possle solutons, Proceedngs of the 5 th Jont Conference on Informton Scences, 999, pp [] L. Polkowsk nd A. Skowron, Towrds dptve clculus of grnules, Proceedngs of 998 IEEE Interntonl Conference on Fuzzy Systems, pp. 6. [3] T.Y. Ln, Grnulr computng on nry reltons I: dt mnng nd neghorhood systems, II: Rough set representtons nd elef functons, n L. Polkowsk nd A. Skowron eds., Rough sets n knowledge dscovery. Hedelerg, Physc-Verlg, 998, pp [4] Y.Y. Yo, Grnulr computng usng neghorhood systems, n R. Roy, T. Furuhsh, nd P.K. Chwdhry (eds.) Advnces n Soft Computng: Engneerng Desgn nd Mnufcturng, Sprnger-Verlg, London, 999, pp [5] T.Y. Ln, Dt mnng: grnulr computng pproch, Proceedngs of the Thrd Pcfc- As Conference on Methodologes for Knowledge Dscovery nd Dt Mnng, 999, pp [6] A. Skowron nd J. Stepnuk, Informton grnules: towrds foundtons of grnulr computng, Interntonl Journl of Intellgent Systems, Vol. 6, 57 85, 00. [7] Y.Y. Yo, Informton grnulton nd rough set pproxmton, Interntonl Journl of Intellgent Systems, Vol. 6, 87 04, 00. [8] J.-S. R. Jng, ANFIS: Adptve-Networksed Fuzzy Inference Systems, IEEE Trnsctons on System, Mn, nd Cyernetcs, vol. 3, no.3, pp , 993. [9] D. C. L, C. Wu, nd F. M. Chng, Usng dt-fuzzfyng technology n smll dt set lernng to mprove FMS schedulng ccurcy, Interntonl Journl of Advnced Mnufcturng Technology, Vol. 7, No. 3-4, pp. 3-38, 005. [0] F. M. Chng, nd C. C. Chn, Improve Neuro- Fuzzy Lernng y Attrute Reducton, The 7 th Annul Meetng of the North Amercn Fuzzy Informton Processng Socety, The Rockefeller Unversty, NY, USA, My 8-, 008. [] B. Predk, R. Slownsk, J. Stefnowsk, R. Susmg, nd Sz. Wlk, ROSE - Softwre Implementton of the Rough Set Theory, In: L. Polkowsk, A. Skowron, eds, Rough Sets nd Current Trends n Computng, Lecture Notes n Artfcl Intellgence, vol. 44, pp , [] B. Predk nd Sz.Wlk, Rough Set Bsed Dt Explorton Usng ROSE System, In: Z. W. Rs, A. Skowron, eds, Foundtons of Intellgent Systems, Lecture Notes n Artfcl Intellgence, Vol. 609, pp.7-80, 999. [3] A. Øhrn nd J. Komorowsk, ROSETTA: rough set toolkt for nlyss of dt, Proc. Thrd Interntonl Jont Conference on Informton Scences, Vol. 3, pp , Durhm, NC, Mrch 997. [4] Z. Pwlk, Rough Sets: Theoretcl Aspects of Resonng out Dt, Kluwer, 99. [5] S. Qng, nd C. Alexos, A modulr pproch to genertng fuzzy rules wth reduced ttrutes for the montorng of complex systems, Engneerng Applctons of Artfcl Intellgence, Vol. 3, No. 3, pp.63-78, 000. [6] M. Beynon, Reducts wthn the vrle precson rough set model: A further nvestgton, Europen Journl of Opertonl Reserch, Vol. 34, pp , 00. [7] J. H. Hsu, T. L. Chng, nd H. C. Wng, VPRS model for mole phone test procedure, Journl of the Chnese Insttute of Industrl Engneers, Vol. 3, no. 4, pp , 006. [8] H. H. Inrn, K. Thngvel, nd A. Pethlkshm, Rough set sed Feture Selecton for We Usge Mnng, Interntonl Conference on Computtonl Intellgence nd Multmed Applctons, pp.33-38, 007. [9] K. K. Ang, nd C. Quek, Stock Trdng Usng RSPOP: A Novel Rough Set-Bsed Neuro- Fuzzy Approch, IEEE Trnsctons on Neurl Network, Vol. 7, no. 5, pp.30-35, 006 [0] Lortory of Intellgent Decson Support Systems, Poznn Unversty of Technology, [] UCI Mchne Lernng Repostory, [] E. J. M. Lurí, J. Duchess, A methodology for developng Byesn networks: An pplcton to nformton technology (IT) mplementton, Europen Journl of ISSN: Issue 7, Volume 8, July 009

11 Fengmng M. Chng Opertonl Reserch Vol. 79, No., pp.34-5, 007. [3] F. M. Chng, The Chrcterstcs of Lernng n Lmted Dt nd the Comprtve Assessment of Lernng Methods, WSEAS Trnsctons on Informton Scence nd Applctons, Vol. 5, No.5, pp , 008. [4] J. R. Qunln, Lernng decson tree clssfers, ACM Computng Surveys Vol. 8, No., pp.7-7, 986. [5] K. Seo, An pplcton of one-clss support vector mchnes n content-sed mge retrevl, Expert Systems Wth Applctons Vol. 33, No., pp , 007. [6] F. M. Chng, Determnton of the Economc Predcton n Smll Dt Set Lernng, WSEAS Trnsctons on Computers, Vol. 5, No., pp , 006. [7] F. M. Chng nd M. Y. Chu, A Method of Smll Dt Set Lernng for Erly Knowledge Acquston, WSEAS Trnsctons on Informton Scence nd Applctons, Vol., No., pp.89-94, 005. [8] F. M. Chng, An ntellgent method for knowledge derved from lmted dt, Proceedngs IEEE Interntonl Conference on Systems, Mn, nd Cyernetcs, Vol., pp [9] D. C. L, C. Wu, T. I. Ts nd F. M. Chng, Usng meg-fuzzfcton nd dt trend estmton n smll dt set lernng for erly FMS schedulng knowledge, Computers & Opertons Reserch, Vol. 33, No.6, pp , 006. [30] D. C. L, C. Wu, nd F. M. Chng, Usng dt contnulzton nd expnson to mprove smll dt set lernng ccurcy for erly flexle mnufcturng system (FMS) schedulng, Interntonl Journl of Producton Reserch, Vol. 44, No., pp , 006. [3] F. M. Chng nd Y. C. Chen, A Frequency Assessment Expert System of Pezoelectrc Trnsducers n Pucty of dt, Expert Systems wth Applctons, Vol. 34, No.4, pp , 008. [3] M. Y. Kng, A comprtve ssessment of clssfcton methods, Decson Support Systems, Vol. 35, pp , 003. ISSN: Issue 7, Volume 8, July 009