A Bayesian Approach toward Active Learning for Collaborative Filtering

Size: px
Start display at page:

Download "A Bayesian Approach toward Active Learning for Collaborative Filtering"

Transcription

1 A Bayesan Appoach towad Actve Leanng fo Collaboatve Flteng Rong Jn Depatment of Compute Scence and Engneeng Mchgan State Unvesty Abstact Collaboatve flteng s a useful technque fo eplotng the pefeence pattens of a goup of uses to pedct the utlty of tems fo the actve use. In geneal, the pefomance of collaboatve flteng depends on the numbe of ated eamples gven by the actve use. The moe the numbe of ated eamples gven by the actve use, the moe accuate the pedcted atngs wll be. Actve leanng povdes an effectve way to acque the most nfomatve ated eamples fom actve uses. Pevous wok on actve leanng fo collaboatve flteng only consdes the epected loss functon based on the estmated model, whch can be msleadng when the estmated model s naccuate. Ths pape takes one step futhe by takng nto account of the posteo dstbuton of the estmated model, whch esults n moe obust actve leanng algothm. Empcal studes wth datasets of move atngs show that when the numbe of atngs fom the actve use s estcted to be small, actve leanng methods only based on the estmated model don t pefom well whle the actve leanng method usng the model dstbuton acheves substantally bette pefomance. 1. Intoducton The apd gowth of the nfomaton on the Intenet demands ntellgent nfomaton agent that can sft though all the avalable nfomaton and fnd out the most valuable to us. Collaboatve flteng eplots the pefeence pattens of a goup of uses to pedct the utlty of tems fo an actve use. Compaed to contentbased flteng appoaches, collaboatve flteng systems have advantages n the envonments whee the contents of tems ae not avalable due to ethe a pvacy ssue o the fact that contents ae dffcult fo a compute to Luo S Language Technoy Insttute School of Compute Scence Canege Mellon Unvesty Pttsbugh, PA ls@cs.cmu.edu analye (e.g. musc and vdeos). One of the key ssues n the collaboatve flteng s to dentfy the goup of uses who shae the smla nteests as the actve use. Usually, the smlaty between uses ae measued based on the atngs ove the same set of tems. Theefoe, to accuately dentfy uses that shae smla nteests as the actve use, a easonably lage numbe of atngs fom the actve use ae usually equed. Howeve, few uses ae wllng to povde atngs fo a lage amount of tems. Actve leanng methods povde a soluton to ths poblem by acqung the atngs fom an actve use that ae most useful n detemnng hs/he nteests. Instead of andomly selectng an tem fo solctng the atng fom the actve use, fo most actve leanng methods, tems ae selected to mame the epected educton n the pedefned loss functon. The commonly used loss functons nclude the entopy of model dstbuton and the pedcton eo. In the pape by Yu et. al. (2003), the epected educton n the entopy of the model dstbuton s used to select the most nfomatve tem fo the actve use. Boutle et. al. (2003) apples the metc of epected value of utlty to fnd the most nfomatve tem fo solctng the atng, whch s to fnd the tem that leads to the most sgnfcant change n the hghest epected atngs. One poblem wth the pevous wok on the actve leanng fo collaboatve flteng s that computaton of epected loss s based only on the estmated model. Ths can be dangeous when the numbe of ated eamples gven by the actve use s small and as a esult the estmated model s usually fa fom beng accuate. A bette stategy fo actve leanng s to take nto account of the model uncetanty by aveagng the epected loss functon ove the posteo dstbuton of models. Wth the full Bayesan teatment, we wll be able to avod the poblem caused by the lage vaance n the model dstbuton. Many studes have been conducted on the actve leanng to take nto account of the model uncetanty. The method of quey by commttee (Seung et. al., 1992; Feud et. al., 1997) smulates the posteo dstbuton of models by constuctng an ensemble of

2 models and the eample wth the lagest uncetanty n pedcton s selected fo use s feedback. In the wok by Tong and Kolle (2000), a full Bayesan analyss of the actve leanng fo paamete estmaton n Bayesan Netwoks s used, whch takes nto account of the model uncetanty n computng the loss functon. In ths pape, we wll apply the full Bayesan analyss to the actve leanng fo collaboatve flteng. Patculaly, n ode to smplfy the computaton, we appomate the posteo dstbuton of model wth a smple Dchlet dstbuton, whch leads to an analytc epesson fo the epected loss functon. The est of the pape s aanged as follows: Secton 2 descbes the elated wok n both collaboatve flteng and actve leanng. Secton 3 dscusses the poposed actve leanng algothm fo collaboatve flteng. The epements ae eplaned and dscussed n Secton 4. Secton 5 concludes ths wok and the futue wok. 2. Related Wok In ths secton, we wll fst befly dscuss the pevous wok on collaboatve flteng, followed by the pevous wok on actve flteng. The pevous wok on actve leanng fo collaboatve flteng wll be dscussed at the end of ths secton. 2.1 Pevous Wok on Collaboatve Flteng Most collaboatve flteng methods fall nto two categoes: Memoy-based algothms and Model-based algothms (Beese et al. 1998). Memoy-based algothms stoe atng eamples of uses n a tanng database. In the pedcatng phase, they pedct the atngs of an actve use based on the coespondng atngs of the uses n the tanng database that ae smla to the actve use. In contast, model-based algothms constuct models that well eplan the atng eamples fom the tanng database and apply the estmated model to pedct the atngs fo actve uses. Both types of appoaches have been shown to be effectve fo collaboatve flteng. In ths subsecton, we focus on the model-based collaboatve flteng appoaches, ncludng the Aspect Model (AM), the Pesonalty Dagnoss (PD) and the Fleble Mtue Model (FMM). Fo the convenence of dscusson, we wll fst ntoduce the annotaton. Let tems denoted by X { 1, 2,..., M }, uses denoted by Y { y1, y2,..., y N }, and the ange of atngs denoted by { 1,..., R }. A tuple (, y, ) means that atng s assgned to tem by use y. Let X ( y) denote the set of tems ated by use y, and R y () stand fo and the atng of tem by use y, espectvely. Aspect model s a pobablstc latent space model, whch models ndvdual pefeences as a conve combnaton of pefeence factos (Hofmann & Pucha 1999; Hofmann, 2003). The latent class vaable Z,,..., } { 1 2 K s assocated wth each pa of a use and an tem. The aspect model assumes that uses and tems ae ndependent fom each othe gven the latent class vaable. Thus, the pobablty fo each obsevaton tuple (, y, ) s calculated as follows: p (, y) p(, ) p( y) (1) Z whee p( y) stands fo the lkelhood fo use y to be n class and p(,) stands fo the lkelhood of assgnng tem wth atng by uses n class. In ode to acheve bette pefomance, the atngs of each use ae nomaled to be a nomal dstbuton wth eo mean and vaance as 1 (Hofmann, 2003). The paamete p(,) s appomated as a Gaussan dstbuton N ( µ, σ ) and p( y) as a multnomal dstbuton. Pesonalty dagnoss appoach teats each use n the tanng database as an ndvdual model. To pedcate the atng of the actve use on cetan tems, we fst compute the lkelhood fo the actve use to be n the model of each tanng use, whch s appomated usng a Gaussan dstbuton: p( y y) e X ( y ) t ( R ( ) R y y 2 ( )) 2σ whee σ stands fo the vaance of Gaussan dstbutons. The atngs fom the tanng use on the same tems ae then weghted by the computed lkelhood. The weghted aveage s used as the estmaton of atngs fo the actve use. Pevous empcal studes have shown that the pesonalty dagnoss method s able to outpefom seveal othe appoaches fo collaboatve flteng (Pennock et al., 2000). Fleble Mtue Model ntoduces two sets of hdden vaables {, y}, wth fo the class of tems and y fo the class of uses (S and Jn, 2003; Jn et. al., 2003). Smla to aspect model, the pobablty fo each obseved tuple (, y, ) s factoed nto a sum ove dffeent classes fo tems and uses,.e., p(, y, ), y p( ) P( ) P( ) P( y ) P(, ) y All the paametes ae estmated usng Epectaton Mamaton algothm (EM) (Dumpste et. al., 1976). The multple-cause vecto quantaton (MCVQ) model (Boutle, Zemel and Maln 2003) uses the smla dea fo collaboatve flteng. 2.2 Pevous Wok on Actve Leanng The goal of actve leanng s to lean the coect model usng only a small numbe of labeled eamples. The geneal appoach s to fnd the eample fom the pool of unlabeled data that gves the lagest educton n the epected loss functon. The loss functons used by most actve leanng methods can be categoed nto two y 2 y (2) (3)

3 goups: the loss functons based on model uncetanty and the loss functon based on pedcton eos. Fo the fst type of loss functons, the goal s to acheve the lagest educton ato n the space of hypothess. One commonly used loss functon s the entopy of the model dstbuton. Methods wthn ths categoy usually select the eample fo whch the model has the lagest uncetanty n pedctng the label (Seung et. al., 1992; Feud et. al, 1997; Abe and Mamtsuka, 1998; Campbell et. al., 2000; Tong and Kolle, 2000). The second type of loss functons nvolved the pedcton eos. The lagest educton n the volume of hypothess space may not necessaly be effectve n cuttng the pedcton eo. Feud et. al. showed an eample n (1997), n whch the lagest educton n the space of hypothess does not bng the optmal mpovement to the educton of pedcton eos. Empcal studes n Bayesan Netwok (Tong and Kolle, 2000) and tet categoaton (Roy and MaCallum, 2001) have shown that usng the loss functon that dectly tagets on the pedcton eo s able to acheve bette pefomance than the loss functon that s only based on the model uncetanty. The commonly used loss functon wthn ths categoy s the entopy of the dstbuton of pedcted labels. In addton to the choce of loss functon, how to estmate the epected loss s anothe mpotant ssue fo actve leanng. Many actve leanng methods compute the epected loss only based on the cuently estmated model wthout takng nto account of the model uncetanty. Even though ths smple stategy woks fne fo many applcatons, t can be vey msleadng, patculaly when the estmated model s fa fom the tue model. As an eample, consdeng leanng a classfcaton model fo the data dstbuton n Fgue 1, whee sphees epesent data ponts of one class and stas epesent data ponts of the othe class. The fou labeled eamples ae hghlghted by the lne-shaded ellpss. Based on these fou tanng eamples, the most lkely decson bounday s the hoontal lne (.e., the dash lne) whle the tue decson bounday s a vetcal lne (.e., the dot lne). If we only ely on the estmated model fo estmatng the epected loss, the eamples that wll be selected fo use s feedback ae most lkely fom the dot-shaded aeas, whch ae neffectve n adustng the estmated decson bounday (.e. the hoontal lne) to the coect decson bounday (.e. the vetcal). On the othe hand, f we can use the model dstbuton fo computng the epected loss, we wll be able to adust the decson bounday moe effectvely snce decson boundaes othe than the estmated one (.e. hoontal lne) ae consdeed n the computaton of epected loss. Thee have been seveal studes on actve leanng that utle the posteo dstbuton of models fo estmatng the epected loss. The quey by commttee appoach smulates the model dstbuton by samplng a set of models out of the posteo dstbuton. In the wok of actve leanng fo paamete estmaton n Bayesan f 1 Estmated Decson Bounday Coect Decson Bounday Fgue 1: A leanng scenao when the estmated model s fa fom the tue model. The vetcal lne coesponds to the coect decson bounday and the hoontal lne coesponds to the estmated decson bounday. The fou labeled eamples ae hghlghted by the lne-shaded aeas. Netwok, a Dchlet dstbuton fo the paametes s used to estmate the change n the entopy functon. Howeve, the geneal dffculty wth the full Bayesan analyss fo actve leanng s the computatonal complety. Fo complcated models, usually t s athe dffcult to obtan the eact posteo dstbuton fo models. As a esult, samplng appoaches such as Makov Chan Mote Calo (MCMC) and Gbbs samplng ae used to appomate the Bayesan aveage. In ths pape, we follow an dea smla to the Laplace appomaton used n gaphc model (MacKay, 1992). Instead of accuately computng the posteos, we appomate the posteo dstbuton wth an analytc epesson and apply the appomated posteo dstbuton to estmate the epected loss. Compaed to the samplng appoaches, ths appoach substantally smplfes the computaton by avodng geneatng a lage numbe of models and calculatng the loss functon values ove the geneated models. 2.3 Pevous Wok on Actve Leanng fo Collaboatve Flteng Thee have been only a few studes on actve leanng fo collaboatve flteng. In the pape by Ka et al (2003), a method smla to Pesonalty Dagnoss (PD) s used fo collaboatve flteng. Each eample s selected fo use s feedback n ode to educe the entopy of the lkemndness dstbuton o p( y y ) n Equaton (2). In the pape by Boutle et al (2003), the Multple-cause vecto quantfcaton method (smla to the FMM model) s used fo collaboatve flteng. Unlke many othe collaboatve flteng eseaches, whch ty to mnme f 2

4 the pedcton eo, ths pape only concens wth the tems that ae stongly ecommended by the system. The loss functon s based on the epected value of nfomaton (EVOI), whch s computed based on the cuently estmated model. One poblem wth both studes on actve leanng fo collaboatve flteng s that, the epected loss s computed only based on a sngle model, namely the cuently estmated model. As llustated by the eample n Fgue 1, the estmaton based on only a sngle model can be msleadng, patculaly when the numbe of ated eamples gven by the actve use s small and meanwhle the numbe of paametes to be estmated s lage. In the late epements, we wll show that the selecton based on the epected loss functon usng only the cuently estmated model can be even wose than smple andom selecton. 3. A Bayesan Appoach Towad Actve Leanng fo Collaboatve Flteng In ths secton, we wll fst dscuss the appomated analytc epesson fo the posteo dstbuton. Then, the appomated posteo dstbuton wll be used to estmate the epected loss. 3.1 Appomate the Model Posteo Dstbuton Fo the sake of smplcty, we wll focus on the actve leanng of aspect model fo collaboatve flteng. Howeve, the pncple used n ths secton can be easly etended to othe models fo collaboatve flteng. As descbed n Equaton (1), each condtonal lkelhood p(, y ) s decomposed as the sum of p (, ) p ( y ). Fo actve use y, the most mpotant task s to detemne ts use type, o p( y ). Let { p( )} stands fo the paamete space. Usually, paametes ae estmated though the mamum lkelhood estmaton,.e. ag ma p(, y ; ) X ( y) ag ma p(, ) X ( y) Z Then, the posteo dstbuton of model,.e., p( D( y )) whee D( y ) ncludes all the atngs gven the actve use y, can be wtten as: (4) whee Z( D( y )) s the nomalaton facto. In above, a unfom dstbuton s used fo po p ( ). Appaently, the posteo dstbuton n (7) s athe dffcult to be used fo estmatng the epected loss due to the multple poducts and the nomalaton facto Z( D( y )). To appomate the posteo dstbuton, we wll consde the epanson of the posteo functon aound the mamum pont. Let stand fo the mamal paametes that ae obtaned fom EM algothm by mamng Equaton (4). Let s consde the ato of p( D( y )) wth espect to p( D( y )), whch can be wtten as: p(, ) p( D( y )) p( D( y )) p(, ) whee p(, ) p(, ) ( α 1) α s defned as p(, ) α + 1 p(, ) (6) (7) Based on the appomaton n Equaton (6), the appomated posteo dstbuton p( D( y )) s a Dchlet wth hype paametes α defned n Equaton (7), o p( D) Γ( α ) α 1 ( α ) Γ whee α α. Futhemoe, the followng elaton between the hype paametes α and optmal paametes s tue: α 1 α 1 Ths s because the optmal paametes obtaned by the EM algothm s actual the fed pont of the followng equaton: (8) (9) 1 p( D( y )) p( D( y ) ) p( ) Z( D( y )) 1 p(, ) Z( D( y )) X ( y) Z (5) p(, ) (10) p(, )

5 Based on the popety n Equaton (9), t s not dffcult to vefy that the optmal pont fo the appomated Dchlet dstbuton n Equaton (8) s eactly. 3.2 Estmate the Epected Loss Smla to many othe actve leanng woks (Tong and Kolle, 2000; Seung et al, 1992; Ka et al, 2003), we use the entopy of the model functon as the loss functon. Many pevous studes on actve leanng ty to fnd the eample that dectly mnmes the entopy of the model paametes wth cuently estmated model. In the case of aspect model, t can be fomulated as the followng optmaton poblem: (11) ag mn,, X p(, ) whee denotes the cuently estmated paametes and denotes the optmal paamete that ae estmated, usng one addtonal ated eample,.e., tem s ated as. The goal of Equaton (11) s to fnd tem such that the epected entopy of dstbuton s mnmed. Thee ae two poblems wth applyng ths smple stategy to collaboatve flteng: 1) The fst poblem comes fom the fact that the epected entopy s computed only based on the cuently estmated model. Accodng to Equaton (11), the epectaton s evaluated ove the dstbuton p(, ). As aleady ponted out n Secton 2, such estmaton could be msleadng patculaly when the cuently estmated model s fa fom the tue model. 2) The second poblem comes fom the chaactestcs of collaboatve flteng. Recall that paamete stands fo the lkelhood fo the actve use to be n the use class. Snce the optmaton n Equaton (11) tes to effcently educe the uncetanty n the type of use class fo the actve use, t wll select eamples that can pufy the class dstbuton. Ths contadcts the pevous studes n collaboatve flteng (S and Jn, 2003), whch found t s moe effectve fo collaboatve flteng to assume a use of multple use types than a sngle type. As a esult of ths obsevaton, we would epect the dstbuton to be of multple modes, not of a sngle model. Based on the above analyss, the optmaton poblem n Equaton (11) s not appopate fo actve leanng of collaboatve flteng. The late empcal studes also ndcate ths fact. In the deal case, f the tue model s gven, the optmal stategy n selectng eample should be: tue tue ag mn (12) X, tue tue p(, ) The goal of Equaton (12) s to fnd an eample such that the updated model paamete, can be adusted tue towad the tue model paamete most effcently. tue Snce the tue model s unknown, we need to appomate the optmaton goal n Equaton (12). One way of appomaton s to eplace the tue model wth the epectaton ove the posteo dstbuton p( D( y )). As a esult of ths appomaton, Equaton (12) s tansfomed nto the followng optmaton poblem: ag mn X, p(, ) p( D( y)) tue (13) Now, the key ssue becomes how to compute the ntegaton effcently snce the samplng methods ae usually tme consumng. Fotunately, the ntegaton n Equaton (13) can be computed analytcally as follows:, p(, ) p( D( y)) p( D( y)) p(, ) d,, p( D( y)) Ψ a + a + ( 2) a Ψ ( a + 1) ( a + 1) a p(, ) a p(, ) a p(, ) a Ψ ( a + 1) p(, ) ( a + 1) a, a ( a + 1) a a p(, ), (, ) a p(, ) a p + a ( a + 1) ( a + 1) a (14) whee Ψ s the dgamma functon. Note that R 1 p(, ) 1 snce a Gaussan appomaton s used to compute p(, ). The above devaton uses the followng popety of dgamma functon: Ψ( α ) Ψ( α ) ~ Dchlet ( α )

6 Wth the analytc esult n Equaton (14), the ntegaton n Equaton (13) can be effcently computed. The othe computatonal complety comes fom the estmaton of the updated model,. The standad way to obtan the updated model s to eun the full EM algothm wth one moe addtonal ated eample (.e., eample s ated as categoy ). Ths can be etemely epensve snce we wll have an updated model fo evey tem and evey possble atng categoy. A moe effcent way fo computng the updated model s to avod the dectly optmaton of Equaton (4). Instead, we use the followng appomaton, ag ma p(, ) p(, y ; ) X ( y ) ag ma p(, ) p( D( y )) (15) ag ma p(, ) Dchlet( α) α 1 ag ma p(, ) The EM updatng equaton fo the above obectve functon s: p(, ) + α 1 α + ( p(, ) 1) (16) The advantage of the updatng equaton n (16) vesus the moe geneal EM updatng equaton n (10) s that t only depends on the atng and tem whle Equaton (10) has to go though all the tems ated by the actve use. In summay, we popose a Bayesan teatment of actve leanng fo collaboatve flteng, whch uses the model posteo dstbuton to compute the estmaton of loss functon. To smplfy the computaton, we appomate the mode dstbuton wth a Dchlet dstbuton. As a esult, the epected loss can be calculated analytcally and the use model can be updated moe effcently. Fo late efeence, we call ths method Bayesan method. 4. Epements In ths secton, we pesent epement esults n ode to addess the followng two questons: 1) Whethe the poposed actve leanng algothm s effectve fo collaboatve flteng? In the epement, we wll compae the poposed algothm to the method of andomly acqung eamples fom the actve use. 2) How mpotant s the full Bayesan teatment? In ths pape, we emphase the mpotance of takng nto account the model uncetanty usng the posteo dstbuton. To llustate ths pont, n ths epement, we Table 1: Chaactestcs of MoveRatng and EachMove. MoveRatng EachMove compae the poposed algothm to two commonly used actve leanng methods that ae only based on the estmated model wthout utlng the model dstbuton. The detals of these two actve leanng methods wll be dscussed late. 4.1 Epement Desgn Two datasets of move atngs wee used n ou epements,.e., MoveRatng 1 and EachMove 2. Fo EachMove, we etacted a subset of 2,000 uses wth moe than 40 atngs. The detals of these two datasets ae lsted n Table 1. Fo MoveRatng dataset, we used the fst 200 uses fo tanng and the est uses fo testng. Fo EachMove dataset, the fst 400 uses wee used fo tanng. Fo each test use, we andomly selected thee tems wth the atngs as the statng seed fo the actve leanng algothm to buld the ntal model. Futhemoe, fo each test use, twenty ated tems wee eseved fo evaluatng the pefomance of dffeent actve leanng methods. Fnally, fo each teaton, an actve leanng algothm s allowed to solct atng fo a sngle tem fom the actve use. Totally, t can ask fo atngs of fve dffeent tems and the pefomance s evaluated fo each feedback. Fo smplcty, we assume that the actve use wll always be able to ate the tems pesented by the actve leanng. Of couse, as ponted out n (Ka et al, 2003), ths s not a completely coect assumpton because thee ae tems that the actve use has not seen befoe and theefoe t s mpossble fo hm/he to ate those tems. Snce the focus of ths pape s on the behavo of actve leanng algothms fo collaboatve flteng, we wll leave ths ssue fo futue wok. The aspect model used fo collaboatve flteng has aleady been descbed n the Secton 2. The numbe of use classes used n the epement s set to be 5 fo MoveRatng dataset and 10 fo EachMove datasets based on the pevous empcal studes. The mean absolute eo (MAE) s used to evaluate the pefomance of collaboatve flteng, whch s defned as follows (Beese et al., 1998): whee LTest Numbe of Uses Numbe of Items Avg. # of ated Items/Use Numbe of Ratngs ^ MAE ( l) Ry ( ( ( ) ) l ) l L (17) Test l s the numbe of the test atngs

7 The poposed algothm s compaed aganst the followng thee actve leanng algothms fo collaboatve flteng: 1) Random Selecton: Ths method andomly selects one tem out of the pool of tems fo use s feedback. We efe ths smple method as andom method. 2) Model Entopy based Sample Selecton. Ths method has aleady been descbed at the begnnng of Secton 3.2. It fnds the tem that effcently educes the entopy of the use class dstbuton fo the actve use (n Equaton (11)). As afoementoned, the poblems wth ths smple selecton stategy ae n two folds: I) Computng the epected loss only based on the cuently estmated model and, II) Conflctng wth the ntuton that each use can be of multple types. By compang ths appoach to the poposed actve leanng method fo collaboatve flteng, we wll be able to see f the dea of usng model dstbuton fo computng epected loss s wothwhle fo collaboatve flteng. We wll efe to ths method as model entopy method. 3) Pedcton Entopy based Sample Selecton. Unlke the pevous method, whch consdes the uncetanty n assgnng the actve use to dffeent use classes, ths method focuses on the uncetanty n pedctng atngs of tems fo the actve use. It selects the tem that s able to educe the entopy of pedcatng atngs fo dffeent tems. Fomally, the selecton cteon can be fomulated as the followng optmaton poblem: ag mn p(,, ) p(,, ) X p(, ) (18) whee notaton, stands fo the updated model usng the eta ated eample (,,y). Smla to the pevous method, ths appoach uses the estmated model fo computng the epected loss. Howeve, unlke the pevous appoach that tes to pufy the dstbuton of use types fo the actve use, ths appoach tagets on the pedcton dstbuton. Theefoe, t s not aganst the ntuton that each use s of multple types. We wll efe ths method as pedcton entopy method. 4.2 Results and Dscusson The esults fo the poposed actve leanng algothm togethe wth the thee baselne models fo database MoveRatng and EachMove ae pesented n Fgue 2 and 3, espectvely. Fst, accodng to Fgue 2 and 3, both the model entopy method and the pedcton entopy method pefom consstently wose than the smple andom method. Ths phenomenon can be eplaned by the fact that the ntal numbe of ated tems gven by the actve use s only thee, whch s elatvely small gven the numbe of paametes to detemne s 5 fo MoveRatng and 10 fo EachMove. As a esult, the epected loss cannot be computed accuately based on the estmated MAE MoveRatng numbe of feedback andom Bayesan model pedcton Fgue 2: MAE esults of fou actve leanng algothms fo collaboatve flteng ove MoveRatng dataset. Legend andom stands fo the andom method, Bayesan fo the Bayesan method, model fo the model entopy method, and pedcaton fo the pedcton entopy method. The smalle the MAE the bette the pefomance Each Move numbe of feedbacks Fgue 3: MAE esults of fou actve leanng algothms fo collaboatve flteng ove EachMove dataset. Legend andom stands fo the andom method, Bayesan fo the Bayesan method, model fo the model entopy method, and pedcaton fo the pedcton entopy method. The smalle the MAE the bette the pefomance. model, and the selected tems wll not be the most nfomatve ones. Futhemoe, compang to the

8 pedcton entopy method, we see that the model entopy method pefoms substantally wose. Fo eample, n Fgue 3, fo the model entopy method, the pefomance of collaboatve flteng s almost unchanged whle the pedcaton entopy method s able to educe the MAE eo fom 1.15 to Ths s because the model entopy method tes to naow down a sngle use type fo the actve use. Snce most uses ae of multple types. t s nappopate to apply the model entopy method to collaboatve flteng. On the othe hand, the pedcton entopy method doesn t have ths defect because t focuses on mnmng the uncetanty n pedcatng atngs of tems fo the actve use nstead of the uncetanty n assgnng use types to the actve use. The second obsevaton fom Fgue 2 and 3 s that the poposed actve leanng method pefoms bette than any of the thee based lne models fo both MoveRatng and EachMove datasets. The most mpotant dffeence between the poposed method and the othe methods fo actve leanng s that the poposed method takes nto account the model dstbuton, whch makes t obust when even thee s only thee ated tems gven by the actve use. 5. Concluson In ths pape, we poposed a full Bayesan teatment of actve leanng fo collaboatve flteng. Dffeent fom pevous studes of actve leanng fo collaboatve flteng, ths method takes nto account the model dstbuton when computng the epected loss. In ode to allevate the computaton complety, we appomate the model posteo wth a smple Dchlet dstbuton. As a esult, the estmated loss can be computed analytcally and the model fo the actve use can be updated effcently. Snce ths wok only focuses on the model qualty, we plan to apply the full Bayesan analyss to the pedcton eo, whch s usually moe effectve accodng to the pevous studes of actve leanng. Refeences Beese, J. S., Heckeman, D., Kade C., (1998). Empcal Analyss of Pedctve Algothms fo Collaboatve Flteng. In the Poceedng of the Fouteenth Confeence on Uncetanty n Atfcal Intellgence. Boutle, C., Zemel, R. S., and Maln, B. (2003). Actve collaboatve flteng. In the Poceedngs of Nneteenth Confeence on Uncetanty n Atfcal Intellgence. Dempste, A. P., Lad, N. M., & Rubn, D. B. (1977). Mamum lkelhood fom ncomplete data va the EM algothm. Jounal of the Royal Statstcal Socety, B39: Feund, Y. Seung, H. S., Sham, E. and Tshby. N. (1997) Selectve samplng usng the quey by commttee algothm. Machne Leanng Hofmann, T., & Pucha, J. (1999). Latent class models fo collaboatve flteng. In Poceedngs of Intenatonal Jont Confeence on Atfcal Intellgence. Hofmann, T. (2003). Gaussan latent semantc models fo collaboatve flteng. In Poceedngs of the 26 th Annual Int l ACM SIGIR Confeence on Reseach and Development n Infomaton Reteval. Jn, R., S, L., Zha., C.X. and Callan J. (2003). Collaboatve flteng wth decoupled models fo pefeences and atngs. In Poceedngs of the 12th Intenatonal Confeence on Infomaton and Knowledge Management. MacKay, D. J. C. (1992). A pactcal Bayesan famewok fo back-popagaton netwoks. Neual Computaton 4(3), Roy, N. and McCallum A. (2001). Towad Optmal Actve Leanng though Samplng Estmaton of Eo Reducton. In Poceedngs of the Eghteenth Intenatonal Confeence on Machne Leanng. Pennock, D. M., Hovt, E., Lawence, S., and Gles, C. L. (2000). Collaboatve flteng by pesonalty dagoss: a hybd memoy- and model-based appoach. In Poceedngs of the Steenth Confeence on Uncetanty n Atfcal Intellgence. Seung, H. S., Oppe, M. and Sompolnsky, H. (1992). Quey by commttee. In Computatonal Leang Theoy, pages S, L. and Jn, R. (2003). Fleble mtue model fo collaboatve flteng. In Poceedngs of the Twenteth Intenatonal Confeence on Machne Leanng. Tong, S. and Kolle, D. (2000). Actve Leanng fo Paamete Estmaton n Bayesan Netwoks. In Advances n Neual Infomaton Pocessng Systems. Yu, K., Schwaghofe, A., Tesp, V., Ma, W.-Y. and Zhang, H. J. (2003). Collaboatve ensemble leanng: combnng collaboatve and content-based Infomaton flteng va heachcal Bayes. In Poceedngs of Nneteenth Intenatonal Confeence on Uncetanty n Atfcal Intellgence. Campbell, C., Cstann, N., and Smola, A. (2000). Quey Leanng wth Lage Magn Classfes. In Poceedngs of 17 th Intenatonal Confeence on Machne Leanng. Abe, N. and H. Mamtsuka (1998). Quey leanng stateges usng boostng and baggng. In Poceedngs of the Ffteenth Intenatonal Confeence on Machne Leanng.

Effectiveness evaluation of electronic warfare command and control system based on grey AHP method

Effectiveness evaluation of electronic warfare command and control system based on grey AHP method Avalable onlne www.jocp.com Jounal of hemcal and Phamaceutcal Reseach, 014, 6(7):535-54 Reseach Atcle ISSN : 0975-7384 ODEN(USA) : JPR5 Effectveness evaluaton of electonc wafae command and contol system

More information

A SAS Macro for Finding Optimal k-means Clustering in One Dimension with Size Constraints

A SAS Macro for Finding Optimal k-means Clustering in One Dimension with Size Constraints Pape SD-02 A SAS Maco fo Fndng Optmal k-means Clusteng n One Dmenson wth Sze Constants Fengao Hu, Geoga Regents Unvesty; Robet E. Johnson, Vandeblt Unvesty ABSTRACT Wang and Song (2011) poposed a k-means

More information

arxiv: v1 [math.co] 21 Jan 2016

arxiv: v1 [math.co] 21 Jan 2016 PROOF OF BIJECTION FOR COMBINATORIAL NUMBER SYSTEM axv:60.05794v [math.co] Jan 06 ABU BAKAR SIDDIQUE, SAADIA FARID, AND MUHAMMAD TAHIR Abstact. Combnatoal numbe system epesents a non-negatve natual numbes

More information

Risk assessment research for power transformer based on fuzzy synthetic evaluation

Risk assessment research for power transformer based on fuzzy synthetic evaluation Avalable onlne www.jocp.com Jounal of Chemcal and Phamaceutcal Reseach, 23, 5(2):62-68 Reseach Atcle ISSN : 975-7384 CODEN(USA) : JCPRC5 Rsk assessment eseach fo powe tansfome based on fuzzy synthetc evaluaton

More information

Planar Curve Representation and Matching

Planar Curve Representation and Matching Plana Cuve Repesentaton and Matchng Mahe Al-Khayat and Fahad Kamanga Compute Scence and Engneeng Depatment Unvesty of Texas at Alngton Alngton, Texas 7609, USA [al-khay kamanga]@cse.uta.edu Abstact In

More information

An Object Based Auto Annotation Image Retrieval System

An Object Based Auto Annotation Image Retrieval System Poceedngs of the 5th WSEAS Intenatonal Confeence on Telecommuncatons and Infomatcs, Istanbul, Tukey, May 27-29, 2006 (pp509-54) An Obect Based Auto Annotaton Image Reteval System Pe-Cheng Cheng, Been-Chan

More information

FUZZY ARTMAP AND NEURAL NETWORK APPROACH TO ONLINE PROCESSING OF INPUTS WITH MISSING VALUES

FUZZY ARTMAP AND NEURAL NETWORK APPROACH TO ONLINE PROCESSING OF INPUTS WITH MISSING VALUES FUZZY ARTMAP AD EURAL ETWORK APPROACH TO OLIE PROCESSIG OF IPUTS WITH MISSIG VALUES F.V. elwamondo* and T. Mawala* * School of Electcal and Infomaton Engneeng, Unvesty of the Wtwatesand, Johannesbug, Pvate

More information

Feature Extraction for Collaborative Filtering: A Genetic Programming Approach

Feature Extraction for Collaborative Filtering: A Genetic Programming Approach IJCSI Intenatonal Jounal of Compute Scence Issues Vol. 9 Issue 5 No Septembe 0 www.ijcsi.og 348 Featue xtacton fo Collaboatve Flteng: A Genetc Pogammng Appoach Deepa Anand Depatment of Compute Scence Chst

More information

One-Dimensional Linear Local Prototypes for Effective Selection of Neuro-Fuzzy Sugeno Model Initial Structure

One-Dimensional Linear Local Prototypes for Effective Selection of Neuro-Fuzzy Sugeno Model Initial Structure One-Dmensonal Lnea Local Pototypes fo Effectve Selecton of euo-fuzzy Sugeno Model Intal Stuctue Jace Kabzńs Insttute of Automatc Contol, Techncal Unvesty of Lodz, Stefanowsego 8/22, 90 924 Lodz, Poland,

More information

Hadoop based Feature Selection and Decision Making Models on Big Data

Hadoop based Feature Selection and Decision Making Models on Big Data Indan Jounal of Scence and Technology, Vol 9(0), DOI: 0.7485/jst/206/v90/88905, Mach 206 ISSN (Pnt) : 0974-6846 ISSN (Onlne) : 0974-5645 Hadoop based Featue Selecton and Decson Makng Models on Bg Data

More information

Journal of Engineering Science and Technology Review 7 (1) (2014) Research of Anti-Noise Image Salient Region Extraction Method

Journal of Engineering Science and Technology Review 7 (1) (2014) Research of Anti-Noise Image Salient Region Extraction Method Jest Jounal of Engneeng Scence and echnology Revew 7 () (04) 43 47 JOURNAL OF Engneeng Scence and echnology Revew www.jest.og Reseach of Ant-Nose Image Salent Regon Extacton Method Bng XU,*, Shy XIE and

More information

FUNDAMENTALS OF RADIOMETRY

FUNDAMENTALS OF RADIOMETRY 1 FUNDAMENTALS OF RADIOMETRY Lectue 5 Radomety 2 Quanttatve measuement of the popetes of EM adaton nteacton wth matte o emsson fom t Radomety deals wth total EM adaton We extend the concept of adans n

More information

Controlled Information Maximization for SOM Knowledge Induced Learning

Controlled Information Maximization for SOM Knowledge Induced Learning 3 Int'l Conf. Atificial Intelligence ICAI'5 Contolled Infomation Maximization fo SOM Knowledge Induced Leaning Ryotao Kamimua IT Education Cente and Gaduate School of Science and Technology, Tokai Univeisity

More information

Robust Proper Clustering Structure Fuzzy Modeling for Function Approximation

Robust Proper Clustering Structure Fuzzy Modeling for Function Approximation Robust Pope lusteng Stuctue Fuzzy Modelng fo Functon Appoxmaton hh-hng Hsao Depatment of Electcal Engneeng Kao Yuan Unvesty Kaohsung ounty,tawan RO hsao@ooneentustedutw Shun-Feng Su Depatment of Electcal

More information

Volumetric Textures. Esteban W. Gonzalez Clua 1 Marcelo Dreux 2

Volumetric Textures. Esteban W. Gonzalez Clua 1 Marcelo Dreux 2 Volumetc Textues Esteban W. Gonzalez Clua 1 Macelo Deux 2 1 PUC-Ro - Depatment of Compute Scence Rua Maquês de São Vcente, 255, 22443-900, Ro de Janeo, RJ, Bazl esteban @nf.puc-o.b 2 PUC-RIO Mechancal

More information

Soochow University: Description and Analysis of the Chinese Word Sense Induction System for CLP2010

Soochow University: Description and Analysis of the Chinese Word Sense Induction System for CLP2010 Soochow Unvesty: Descpton and Analyss of the Chnese Wod Sense Inducton System fo CLP2010 Hua Xu Bng Lu Longhua Qan Guodong Zhou Natual Language Pocessng Lab School of Compute Scence and Technology Soochow

More information

Fuzzy Membership Function Based on Structural Information of Data Fang-Zhi ZHU a, Hui-Ru WANG b, Zhi-Jian ZHOU c,*

Fuzzy Membership Function Based on Structural Information of Data Fang-Zhi ZHU a, Hui-Ru WANG b, Zhi-Jian ZHOU c,* 06 Intenatona Confeence on Sevce Scence, Technoogy and Engneeng (SSTE 06 ISBN: 978--60595-35-9 Fuzzy Membeshp Functon Based on Stuctua Infomaton of Data Fang-Zh ZHU a, Hu-Ru WANG b, Zh-Jan ZHOU c,* Coege

More information

Fuzzy clustering ensemble based on mutual information

Fuzzy clustering ensemble based on mutual information Poceedngs of the 6th WSEAS Intenatonal Confeence on Smulaton, Modellng and Optmzaton, Lsbon, Potugal, Septembe 22-24, 26 476 Fuzz clusteng ensemble based on mutual nfomaton YAN GAO SHIWEN GU LIMING XIA

More information

Fault Tolerant Routing For Wireless Sensor Grid Networks

Fault Tolerant Routing For Wireless Sensor Grid Networks Fault oleant Routng Fo Weless Senso Gd etwoks Xn-ng Huang, Jng Deng 2, Jng a and Zeyu Wu Depatment of Electcal Engneeng 2 Depatment of Compute Scence Unvesty of ew Oleans, ew Oleans, A 7048, USA Abstact

More information

Comparing performance of SRAMT-LE vs other layered encoding schemes regarding TCP friendliness

Comparing performance of SRAMT-LE vs other layered encoding schemes regarding TCP friendliness Compang pefomance of SRAMT-LE vs othe layeed encodng schemes egadng TCP fendlness Chstos Bouas, Apostolos Gkamas Reseach Academc Compute Technology Insttute, 61 Rga Feaou St, GR-26221 Patas, Geece Compute

More information

Supervised and Unsupervised Text Classification via Generic Summarization

Supervised and Unsupervised Text Classification via Generic Summarization Intenatonal Jounal of Compute Infomaton Systems and Industal Management Applcatons. ISSN 50-7988 Volume 5 (03) pp. 509-55 MIR Labs, www.mlabs.net/jcsm/ndex.html Supevsed and Unsupevsed Text Classfcaton

More information

NODAL AND LOOP ANALYSIS TECHNIQUES. Develop systematic techniques to determine all the voltages and currents in a circuit NODE ANALYSIS

NODAL AND LOOP ANALYSIS TECHNIQUES. Develop systematic techniques to determine all the voltages and currents in a circuit NODE ANALYSIS NODAL AND LOOP ANALYSS TECHNQUES LEANNG GOALS NODAL ANALYSS LOOP ANALYSS Develop systematic techniques to determine all the voltages and currents in a circuit NODE ANALYSS One of the systematic ways to

More information

Dynamic Optimization of Structures Subjected to Earthquake

Dynamic Optimization of Structures Subjected to Earthquake IACSIT Intenatonal Jounal of Engneeng and Technology, Vol. 8, No. 4, August 06 Dynamc Optmzaton of Stuctues Subjected to Eathquake Aleza Lavae and Aleza Lohasb Abstact To educe the oveall tme of stuctual

More information

Enhance the Alignment Accuracy of Active Shape Models Using Elastic Graph Matching

Enhance the Alignment Accuracy of Active Shape Models Using Elastic Graph Matching Enhance the Algnment Accuacy of Actve Shape Models Usng Elastc Gaph Matchng Sanqang Zhao 1,, Wen Gao, Shguang Shan, Baoca Yn 1 1) Multmeda and Intellgent Softwae Technology Lab, Beng Unvesty of Technology,

More information

Constructing Service Semantic Link Network Based on the Probabilistic Graphical Model

Constructing Service Semantic Link Network Based on the Probabilistic Graphical Model Intenatonal Jounal of omputatonal Intellgence ystems, Vol. 5, No. 6 (Novembe, 2012), 1040-1051 onstuctng evce emantc Lnk Netwok Based on the Pobablstc Gaphcal odel ANPING ZHAO * ollege of ompute and Infomaton

More information

Performance Evaluation of Neighborhood Signature Techniques for Peer-to-Peer Search

Performance Evaluation of Neighborhood Signature Techniques for Peer-to-Peer Search Pefomance Evaluaton of Neghbohood gnatue Technques fo Pee-to-Pee each e L Wang-Chen Lee * Anand vasubamanam Depatment of Compute cence and Engneeng Pennsylvana tate Unvesty Unvesty Pak, PA 16802, UA E-al:

More information

A QoS Scheme for a Congestion Core Network Based on Dissimilar QoS Structures in Smart-Phone Environments

A QoS Scheme for a Congestion Core Network Based on Dissimilar QoS Structures in Smart-Phone Environments Sensos 2010, 10, 10006-10013; do:10.3390/s101110006 OPEN ACCESS sensos ISSN 1424-8220 www.mdp.com/jounal/sensos Atcle A QoS Scheme fo a Congeston Coe Netwok Based on Dssmla QoS Stuctues n Smat-Phone Envonments

More information

DATA DIMENSIONALITY REDUCTION METHODS FOR ORDINAL DATA

DATA DIMENSIONALITY REDUCTION METHODS FOR ORDINAL DATA DATA DIMENIONALITY REDUCTION METHOD FOR ORDINAL DATA Matn Pokop Hana Řezanková Abstact Fom questonnae suvey we fequently get data, the values ae expessed n odnal (e.g. Lket) scale. The questonnae contans

More information

SOFT COMPUTING OPTIMIZATION TECHNIQUES FOR SOLAR PHOTOVOLTAIC ARRAYS

SOFT COMPUTING OPTIMIZATION TECHNIQUES FOR SOLAR PHOTOVOLTAIC ARRAYS ARPN Jounal of Engneeng and Appled Scences 2006-20 Asan Reseach Publshng Netwok (ARPN). All ghts eseved. SOFT COMPUTING OPTIMIZATION TECHNIQUES FOR SOLAR PHOTOVOLTAIC ARRAYS Ramapabha R. and Mathu B. L.

More information

Outlier Detection in 3D Coordinate Transformation with Fuzzy Logic

Outlier Detection in 3D Coordinate Transformation with Fuzzy Logic Acta Montanstca Slovaca Ročník 7 (22), číslo, -8 Outle Detecton n 3D Coodnate ansfomaton wth Fuzzy Logc Yasemn Ssman, Aslan Dlave 2 and Sebahattn Bektas Coodnate measuements nevtably contan outles that

More information

Recognition of Shapes for Object Retrieval in Image Databases by Discrete Curve Evolution and Two Consecutive Primitive Edges

Recognition of Shapes for Object Retrieval in Image Databases by Discrete Curve Evolution and Two Consecutive Primitive Edges Poceedngs of the Intenatonal MultConfeence of Engnees and Compute Scentsts 009 Vol I IMECS 009 Mach 8-0 009 Hong Kong Recognton of Shapes fo Object Reteval n Image Databases by Dscete Cuve Evoluton and

More information

ARTICULATED MOTION ANALYSIS FROM MOTION CAPTURE DATA

ARTICULATED MOTION ANALYSIS FROM MOTION CAPTURE DATA ARICULAED MOION ANALYSIS FROM MOION CAPURE DAA J. Fayad, A. Del Bue and P. M. Q. Agua 3. ABSRAC We pesent computatonal methods to extact and model dffeent onts of a genec subect, n an automatc way. he

More information

An Energy-Aware QoS Routing Protocol for Wireless Sensor Networks

An Energy-Aware QoS Routing Protocol for Wireless Sensor Networks An Enegy-Awae QoS Routng Potocol fo Weless Senso Netwoks Kemal Akkaya and Mohamed Youns Depatment of Compute Scence and Electcal Engneeng Unvesty of Mayland, Baltmoe County Baltmoe, MD 225 kemal youns@cs.umbc.edu

More information

An Optimization Procedure In A Production Line Of Sokat Soap Industry, Ikotun, Lagos State. Adamawa State, Nigeria.

An Optimization Procedure In A Production Line Of Sokat Soap Industry, Ikotun, Lagos State. Adamawa State, Nigeria. An Optmzaton Pocedue In A Poducton Lne Of Sokat Soap Industy, Ikotun, Lagos State. 1 O.S. Balogun, 2 A.A. Olatun and 3 A.A. Momoh 1 Depatment of Statstcs and Opeatons Reseach, Modbbo Adama Unvesty of Technology,

More information

IP Network Design by Modified Branch Exchange Method

IP Network Design by Modified Branch Exchange Method Received: June 7, 207 98 IP Netwok Design by Modified Banch Method Kaiat Jaoenat Natchamol Sichumoenattana 2* Faculty of Engineeing at Kamphaeng Saen, Kasetsat Univesity, Thailand 2 Faculty of Management

More information

A Neural Network Model for Storing and Retrieving 2D Images of Rotated 3D Object Using Principal Components

A Neural Network Model for Storing and Retrieving 2D Images of Rotated 3D Object Using Principal Components A Neual Netwok Model fo Stong and Reteving 2D Images of Rotated 3D Object Using Pncipal Components Tsukasa AMANO, Shuichi KUROGI, Ayako EGUCHI, Takeshi NISHIDA, Yasuhio FUCHIKAWA Depatment of Contol Engineeng,

More information

Performance Evaluation of WDM LAN s with Propagation Delay Constraint Analysis

Performance Evaluation of WDM LAN s with Propagation Delay Constraint Analysis Pefomance Evaluaton of WD AN s wth Popagaton Dela Constant Analss IOANNI E. POUNOURAKI Depatment of Electcal & Compute Engneeng Natonal echncal Unvest of Athens 57 73 Zogaphou, Athens GREECE Abstact: -In

More information

Vol. 4, No. 6 June 2013 ISSN Journal of Emerging Trends in Computing and Information Sciences CIS Journal. All rights reserved.

Vol. 4, No. 6 June 2013 ISSN Journal of Emerging Trends in Computing and Information Sciences CIS Journal. All rights reserved. Vol 4, No 6 June 03 ISSN 079-8407 Jounal Emegng Tends n Computng and Infomaton Scences 009-03 CIS Jounal All ghts eseved http://wwwcsjounalog Optmal Powe Dstbuton System Placement to Install the RTU wth

More information

An Unsupervised Segmentation Framework For Texture Image Queries

An Unsupervised Segmentation Framework For Texture Image Queries An Unsupevised Segmentation Famewok Fo Textue Image Queies Shu-Ching Chen Distibuted Multimedia Infomation System Laboatoy School of Compute Science Floida Intenational Univesity Miami, FL 33199, USA chens@cs.fiu.edu

More information

Detection and Recognition of Alert Traffic Signs

Detection and Recognition of Alert Traffic Signs Detection and Recognition of Alet Taffic Signs Chia-Hsiung Chen, Macus Chen, and Tianshi Gao 1 Stanfod Univesity Stanfod, CA 9305 {echchen, macuscc, tianshig}@stanfod.edu Abstact Taffic signs povide dives

More information

Interactive NURBS Tutorial in Virtual Reality

Interactive NURBS Tutorial in Virtual Reality Inteacte NURBS Ttoal n Vtal Realty Eng 477 Poect Fa Octobe 10 2006 Edmond Nadle enadle@mch.ed 734 913-6970 Eng 477 Poect Fa Octobe 10 2006 Eng 477 Poect Fa Octobe 10 2006 NURBS Ces and Sfaces = Non-Unfom

More information

Improved Methods on PCA Based Human Face Recognition for Distorted Images

Improved Methods on PCA Based Human Face Recognition for Distorted Images Poceedngs of the Intenatonal MultConfeence of Engnees and Compute Scentsts 016 Vol I,, Mach 16-18, 016, Hong Kong Impoved Methods on PCA Based Human Face Recognton fo Dstoted Images Buce Poon, M. Ashaful

More information

AN EXPLICIT GROWTH MODEL OF THE STEREO REGION GROWING ALGORITHM FOR PARALLEL PROCESSING

AN EXPLICIT GROWTH MODEL OF THE STEREO REGION GROWING ALGORITHM FOR PARALLEL PROCESSING Intenatonal Achves of Photogammety, Remote Sensng and Spatal Infomaton Scences, Vol. XXXVIII, Pat 5 AN EXPLICIT GROWTH MODEL OF THE STEREO REGION GROWING ALGORITHM FOR PARALLEL PROCESSING Dongoe Shn*,

More information

COVERAGE BASED DENSITY ESTIMATION STRATEGY FOR VANET

COVERAGE BASED DENSITY ESTIMATION STRATEGY FOR VANET Intenatonal Jounal of Advanced Compute Technology (IJACT) ISSN:319-7900 COVERAGE BASED DENSITY ESTIMATION STRATEGY FOR VANET Chun-Chh Lo, Depatment of Compute Scence and Infomaton Engneeng, Natonal Cheng

More information

Recognizing Facial Expressions Based on Gabor Filter Selection

Recognizing Facial Expressions Based on Gabor Filter Selection Recognzng Facal Expessons Based on Gabo Flte Selecton Zang Zhang, Xaomn Mu, Le Gao School of Infomaton Engneeng Zhengzhou Unvest Zhengzhou, Chna Abstact Recognton of human emotonal state s an mpotant component

More information

KANBAN DEVS MODELLING, SIMULATION AND VERIFICATION

KANBAN DEVS MODELLING, SIMULATION AND VERIFICATION KANBAN DEVS MODELLING, SIMULATION AND VERIFICATION Alex Cave and Saed Nahavand School of Engneeng and Technology, Deakn Unvesty, Geelong, Vctoa 3216, Austala nahavand@deakn.edu.au ABSTRACT Kanban Contol

More information

Weighted Piecewise LDA for Solving the Small Sample Size Problem in Face Verification

Weighted Piecewise LDA for Solving the Small Sample Size Problem in Face Verification TNN05-P485 Weghted Pecewse LDA fo Solvng the Small Sample Sze Poblem n Face Vefcaton Maos Kypeountas, Anastasos Tefas, Membe, IEEE, and Ioanns Ptas, Seno Membe, IEEE Abstact A novel algothm that can be

More information

Clustering Interval-valued Data Using an Overlapped Interval Divergence

Clustering Interval-valued Data Using an Overlapped Interval Divergence Poc. of the 8th Austalasian Data Mining Confeence (AusDM'9) Clusteing Inteval-valued Data Using an Ovelapped Inteval Divegence Yongli Ren Yu-Hsn Liu Jia Rong Robet Dew School of Infomation Engineeing,

More information

THE MODELLING OF INVOLUTE TEETH GENERATION, WITH THE RELATIVE GENERATING TRAJECTORIES METHOD

THE MODELLING OF INVOLUTE TEETH GENERATION, WITH THE RELATIVE GENERATING TRAJECTORIES METHOD THE MODELLING OF INVOLUTE TEETH GENERATION, WITH THE RELATIVE GENERATING TRAJECTORIES METHOD Vgl Gabel TEODOR 1, Ncuşo BAROIU 1, Flon SUSAC 1 and Ncolae OANCEA 1 ABSTRACT: The modellng of a cul of sufaces

More information

Adjusting the Contact Surface of Forming Tools in Order to Compensate for Elastic Deformations during the Process

Adjusting the Contact Surface of Forming Tools in Order to Compensate for Elastic Deformations during the Process Adjustng the Contact Suface of Fomng Tools n Ode to Compensate fo Elastc Defomatons dung the Pocess Knut Goßmann, Hajo Weme, Andè Hadtmann, Las Pente, Sebastan Kechenbaue Insttute fo Machne Tools and Contol

More information

Point-Biserial Correlation Analysis of Fuzzy Attributes

Point-Biserial Correlation Analysis of Fuzzy Attributes Appl Math Inf Sci 6 No S pp 439S-444S (0 Applied Mathematics & Infomation Sciences An Intenational Jounal @ 0 NSP Natual Sciences Publishing o Point-iseial oelation Analysis of Fuzzy Attibutes Hao-En hueh

More information

Statistical Bayesian Learning for Automatic Arabic Text Categorization

Statistical Bayesian Learning for Automatic Arabic Text Categorization Jounal of Compute Scence 7 (): 39-45, 20 ISSN 549-3636 20 Scence Publcatons Statstcal Baesan Leanng fo Automatc Aabc Text Categozaton Bassam Al-Salem and Mohd. Juzaddn Ab Azz Depatment of Compute Scence,

More information

An Information Maximization Model of Eye Movements

An Information Maximization Model of Eye Movements An Infomaton Maxmzaton Model of Eye Movements Laua Walke Rennnge, James Coughlan, Peet Veghese Smth-Kettlewell Eye Reseach Insttute {laua, coughlan, peet}@sk.og Jtenda Malk Unvesty of Calfona, Bekeley

More information

Journal of World s Electrical Engineering and Technology J. World. Elect. Eng. Tech. 1(1): 12-16, 2012

Journal of World s Electrical Engineering and Technology J. World. Elect. Eng. Tech. 1(1): 12-16, 2012 2011, Scienceline Publication www.science-line.com Jounal of Wold s Electical Engineeing and Technology J. Wold. Elect. Eng. Tech. 1(1): 12-16, 2012 JWEET An Efficient Algoithm fo Lip Segmentation in Colo

More information

A LIGHT SCATTERING METHOD TO MEASURE REAL-TIME PARTICULATE EMISSIONS

A LIGHT SCATTERING METHOD TO MEASURE REAL-TIME PARTICULATE EMISSIONS A Lght Scatteng Method to Measue Real-tme Patculate Emssons Rose Gong & Keyn Wllams 6 th Austalasan Tanspot Reseach Foum Wellngton New Zealand -3 Octobe 003 A LIGHT SCATTERING METHOD TO MEASURE REAL-TIME

More information

Positioning of a robot based on binocular vision for hand / foot fusion Long Han

Positioning of a robot based on binocular vision for hand / foot fusion Long Han 2nd Intenational Confeence on Advances in Mechanical Engineeing and Industial Infomatics (AMEII 26) Positioning of a obot based on binocula vision fo hand / foot fusion Long Han Compute Science and Technology,

More information

Effective Missing Data Prediction for Collaborative Filtering

Effective Missing Data Prediction for Collaborative Filtering Effective Missing Data Pediction fo Collaboative Filteing Hao Ma, Iwin King and Michael R. Lyu Dept. of Compute Science and Engineeing The Chinese Univesity of Hong Kong Shatin, N.T., Hong Kong { hma,

More information

An Extension to the Local Binary Patterns for Image Retrieval

An Extension to the Local Binary Patterns for Image Retrieval , pp.81-85 http://x.oi.og/10.14257/astl.2014.45.16 An Extension to the Local Binay Pattens fo Image Retieval Zhize Wu, Yu Xia, Shouhong Wan School of Compute Science an Technology, Univesity of Science

More information

Real Time Face Recognition Using Polynomial Regression and Sub-region Color Component Distribution

Real Time Face Recognition Using Polynomial Regression and Sub-region Color Component Distribution Real Tme Face Recognton Usng Polynomal Regesson and Sub-egon Colo Component Dstbuton Manshanka Mondal, Md. Almas Hossan, Md. Matu Rahman, Kamul Hasan Talukde* Compute Scence and Engneeng Dscplne Khulna

More information

Feature Reduction and Selection

Feature Reduction and Selection Feature Reducton and Selecton Dr. Shuang LIANG School of Software Engneerng TongJ Unversty Fall, 2012 Today s Topcs Introducton Problems of Dmensonalty Feature Reducton Statstc methods Prncpal Components

More information

Rigid Body Segmentation and Shape Description from Dense Optical Flow Under Weak Perspective

Rigid Body Segmentation and Shape Description from Dense Optical Flow Under Weak Perspective , VOL. 19, NO. 2, FEBRUARY 1997 139 Rgd Body Segmentaton and Shape Descpton fom Dense Optcal Flow Unde Weak Pespectve Joseph Webe and Jtenda Malk Abstact We pesent an algothm fo dentfyng and tackng ndependently

More information

School of Computer Science. control algorithms for use in datagram networks. rates to alleviate the congestion. In the case

School of Computer Science. control algorithms for use in datagram networks. rates to alleviate the congestion. In the case WF 2 Q: Wost-case Fa Weghted Fa Queueng Jon C.R. Bennett FORE Systems Hu Zhang School of Compute Scence Canege Mellon Unvesty Abstact The Genealzed Pocesso Shang (GPS) dscplne s poven to have two desable

More information

A Two-stage and Parameter-free Binarization Method for Degraded Document Images

A Two-stage and Parameter-free Binarization Method for Degraded Document Images A Two-stage and Paamete-fee Binaization Method fo Degaded Document Images Yung-Hsiang Chiu 1, Kuo-Liang Chung 1, Yong-Huai Huang 2, Wei-Ning Yang 3, Chi-Huang Liao 4 1 Depatment of Compute Science and

More information

HISTOGRAMS are an important statistic reflecting the

HISTOGRAMS are an important statistic reflecting the JOURNAL OF L A T E X CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 D 2 HistoSketch: Disciminative and Dynamic Similaity-Peseving Sketching of Steaming Histogams Dingqi Yang, Bin Li, Laua Rettig, and Philippe

More information

Flood Runoff Simulation Using Grid-based Radar Rainfall and Vflo TM Model

Flood Runoff Simulation Using Grid-based Radar Rainfall and Vflo TM Model Flood Runoff Smulaton Usng Gd-based Rada Ranfall and Vflo TM Model HUI SEONG NOH 1, BYUNG SIK KIM 2, NA RAE KANG 3, HUNG SOO KIM 4 1 Dept. of Cvl Engneeng, Inha Unvesty, Incheon, Koea, heesung80@hanmal.net

More information

COMPARISON OF DIMENSIONALITY REDUCTION METHODS APPLIED TO ORDINAL DATA

COMPARISON OF DIMENSIONALITY REDUCTION METHODS APPLIED TO ORDINAL DATA The 7 th Intenatonal Days of tatstcs and Economcs, Pague, eptembe 19-21, 2013 COMPARION OF DIMENIONALITY REDUCTION METHOD APPLIED TO ORDINAL DATA Matn Pokop Hana Řezanková Abstact The pape deals wth the

More information

NURBS Curve Shape Modification and Fairness Evaluation for Computer Aided Aesthetic Design

NURBS Curve Shape Modification and Fairness Evaluation for Computer Aided Aesthetic Design oceedngs of the 7th WSEAS Int. Conf. on Sgnal ocessng, Computatonal Geomet & Atfcal Vson, Athens, Geece, August 4-6, 7 38 NURBS Cuve Shape Modfcaton and Faness Evaluaton fo Compute Aded Aesthetc Desgn

More information

Optical Flow for Large Motion Using Gradient Technique

Optical Flow for Large Motion Using Gradient Technique SERBIAN JOURNAL OF ELECTRICAL ENGINEERING Vol. 3, No. 1, June 2006, 103-113 Optical Flow fo Lage Motion Using Gadient Technique Md. Moshaof Hossain Sake 1, Kamal Bechkoum 2, K.K. Islam 1 Abstact: In this

More information

Research on Fast Reconstruction of Virtual Plant Leaves

Research on Fast Reconstruction of Virtual Plant Leaves Reseach on Fast Reconstucton of Vtual Plant Leaves Fang Ku Lu-Mng Shen 2 College of Infomaton Scence and Technology Hunan Agcultual Unvesty, Changsha, Chna fk@hunau.net Jng Song 2 College of Scence Hunan

More information

Manuscript Draft. Title: Lattice Boltzmann Model for Fast Level Set Algorithm Using the Multiple Kernel Fuzzy C-Means

Manuscript Draft. Title: Lattice Boltzmann Model for Fast Level Set Algorithm Using the Multiple Kernel Fuzzy C-Means Undestandng Compute Vson and Image Manuscpt Daft Manuscpt Numbe: CVIU-12-298 Ttle: Lattce Boltzmann Model fo Fast Level Set Algothm Usng the Multple Kenel Fuzzy C-Means Atcle Type: Regula Pape Keywods:

More information

Module 6 STILL IMAGE COMPRESSION STANDARDS

Module 6 STILL IMAGE COMPRESSION STANDARDS Module 6 STILL IMAE COMPRESSION STANDARDS Lesson 17 JPE-2000 Achitectue and Featues Instuctional Objectives At the end of this lesson, the students should be able to: 1. State the shotcomings of JPE standad.

More information

Interactive Rendering of Deforming NURBS Surfaces

Interactive Rendering of Deforming NURBS Surfaces EUROGRAHICS 97 / D. Fellne and L. Szmay-Kalos Volume 6, (997), Numbe 3 (Guest Edtos) Inteactve Rendeng of Defomng NURBS Sufaces Fedeck W. B. L Rynson W. H. Lau Mak Geen Compute Gaphcs and Meda Laboatoy

More information

Some Image Processing Algorithms on a RAP with Wider Bus Networks

Some Image Processing Algorithms on a RAP with Wider Bus Networks Some Image Pocessng Algothms on a RAP wth Wde Bus Netwoks Shung-Shng Lee, Sh-Jnn Hong, Hong-Ren Tsa and Yu-Hua Lee Natonal Tawan Insttute of Technology Depatment of Electcal Engneeng 4, Secton 4, Kee-Lung

More information

A METHODOLOGY FOR RATING AND RANKING HAZARDS IN MARITIME FORMAL SAFETY ASSESSMENT USING FUZZY

A METHODOLOGY FOR RATING AND RANKING HAZARDS IN MARITIME FORMAL SAFETY ASSESSMENT USING FUZZY Doumas N. Geogos, Nktakos V. Nñtas, ambou A. aa A ETODOOGY FOR RATING AND RANKING AZARDS IN ARITIE FORA SAFETY ASSESSENT USING FUZZY OGIC A ETODOOGY FOR RATING AND RANKING AZARDS IN ARITIE FORA SAFETY

More information

Multi-Objective Optimization Using LS-OPT

Multi-Objective Optimization Using LS-OPT 6. LS-DYNA Anwendefoum, fankenthal 007 Optmeung Mult-Objectve Optmzaton Usng LS-OPT Tusha Goel, Nelen Stande Lvemoe Softwae Technology Copoaton, Lvemoe, CA, USA Summay: Most engneeng optmzaton poblems

More information

3D Head Tracking Based on Recognition and Interpolation Using a Time-Of- Flight Depth Sensor

3D Head Tracking Based on Recognition and Interpolation Using a Time-Of- Flight Depth Sensor 3D Head Tacng Based on Recognton and Intepolaton Usng a Tme-Of- Flght Depth Senso Salh Bua Götü 1 and Calo Tomas 1,2 1 Canesta Inc., 2 Due Unvesty bgotu@canesta.co tomas@cs.due.edu Abstact Ths pape descbes

More information

y and the total sum of

y and the total sum of Lnear regresson Testng for non-lnearty In analytcal chemstry, lnear regresson s commonly used n the constructon of calbraton functons requred for analytcal technques such as gas chromatography, atomc absorpton

More information

GENERAL MODEL FOR AIRBORNE AND SPACEBORNE LINEAR ARRAY SENSORS

GENERAL MODEL FOR AIRBORNE AND SPACEBORNE LINEAR ARRAY SENSORS GENERAL MODEL FOR AIRBORNE AND SPAEBORNE LINEAR ARRAY SENSORS Danela Pol Insttute of Geodesy and Photogammety, Swss Fedeal Insttute of Technology, Zuch, Swtzeland danela@geod.baug.ethz.ch KEY WORDS: Oentaton,

More information

Gravitational Shift for Beginners

Gravitational Shift for Beginners Gavitational Shift fo Beginnes This pape, which I wote in 26, fomulates the equations fo gavitational shifts fom the elativistic famewok of special elativity. Fist I deive the fomulas fo the gavitational

More information

Fair NURBS Curve Generation using a Hand-drawn Sketch for Computer Aided Aesthetic Design

Fair NURBS Curve Generation using a Hand-drawn Sketch for Computer Aided Aesthetic Design Poceedngs of the 7th WSEAS Int. Conf. on Sgnal Pocessng, Computatonal Geomet & Atfcal Vson, Athens, Geece, August 24-26, 27 29 Fa NURBS Cuve Geneaton usng a Hand-dawn Sketch fo Compute Aded Aesthetc Desgn

More information

Meta-heuristics for Multidimensional Knapsack Problems

Meta-heuristics for Multidimensional Knapsack Problems 2012 4th Internatonal Conference on Computer Research and Development IPCSIT vol.39 (2012) (2012) IACSIT Press, Sngapore Meta-heurstcs for Multdmensonal Knapsack Problems Zhbao Man + Computer Scence Department,

More information

Query Language #1/3: Relational Algebra Pure, Procedural, and Set-oriented

Query Language #1/3: Relational Algebra Pure, Procedural, and Set-oriented Quey Language #1/3: Relational Algeba Pue, Pocedual, and Set-oiented To expess a quey, we use a set of opeations. Each opeation takes one o moe elations as input paamete (set-oiented). Since each opeation

More information

Effects of Model Complexity on Generalization Performance of Convolutional Neural Networks

Effects of Model Complexity on Generalization Performance of Convolutional Neural Networks Effects of Model Complexity on Genealization Pefomance of Convolutional Neual Netwoks Tae-Jun Kim 1, Dongsu Zhang 2, and Joon Shik Kim 3 1 Seoul National Univesity, Seoul 151-742, Koea, E-mail: tjkim@bi.snu.ac.k

More information

Graph Cut-based Automatic Segmentation of Lung Nodules using Shape, Intensity, and Spatial Features

Graph Cut-based Automatic Segmentation of Lung Nodules using Shape, Intensity, and Spatial Features Second Intenatonal Wokshop on -103- Gaph Cut-based Automatc Segmentaton of Lung Nodules ung Shape, Intenty, and Spatal eatues Xujong Ye, Gaeth Beddoe and Geg Slabaugh Medcght PLC, London, UK Abstact. Ths

More information

Towards Adaptive Information Merging Using Selected XML Fragments

Towards Adaptive Information Merging Using Selected XML Fragments Towads Adaptive Infomation Meging Using Selected XML Fagments Ho-Lam Lau and Wilfed Ng Depatment of Compute Science and Engineeing, The Hong Kong Univesity of Science and Technology, Hong Kong {lauhl,

More information

A modal estimation based multitype sensor placement method

A modal estimation based multitype sensor placement method A modal estimation based multitype senso placement method *Xue-Yang Pei 1), Ting-Hua Yi 2) and Hong-Nan Li 3) 1),)2),3) School of Civil Engineeing, Dalian Univesity of Technology, Dalian 116023, China;

More information

Spiral Recognition Methodology and Its Application for Recognition of Chinese Bank Checks

Spiral Recognition Methodology and Its Application for Recognition of Chinese Bank Checks Spial Recognition Methodology and Its Application fo Recognition of Chinese Bank Checks Hanshen Tang 1, Emmanuel Augustin 2, Ching Y. Suen 1, Olivie Baet 2, Mohamed Cheiet 3 1 Cente fo Patten Recognition

More information

THE HIERARCHICAL MODEL OF INTERACTION BETWEEN INTELLIGENT AGENTS IN THE MANET CONTROL SYSTEMS

THE HIERARCHICAL MODEL OF INTERACTION BETWEEN INTELLIGENT AGENTS IN THE MANET CONTROL SYSTEMS UDC 6.396.4 THE HIERARCHICAL MODEL OF INTERACTION BETWEEN INTELLIGENT AGENTS IN THE MANET CONTROL SYSTEMS Oleg Ya. Sova Valey A. Romanyuk Mltay Insttute of Telecommuncatons and Infomatzaton Kyv Ukane Dmyto

More information

Generalized Grey Target Decision Method Based on Decision Makers Indifference Attribute Value Preferences

Generalized Grey Target Decision Method Based on Decision Makers Indifference Attribute Value Preferences Ameican Jounal of ata ining and Knowledge iscovey 27; 2(4): 2-8 http://www.sciencepublishinggoup.com//admkd doi:.648/.admkd.2724.2 Genealized Gey Taget ecision ethod Based on ecision akes Indiffeence Attibute

More information

Input Layer f = 2 f = 0 f = f = 3 1,16 1,1 1,2 1,3 2, ,2 3,3 3,16. f = 1. f = Output Layer

Input Layer f = 2 f = 0 f = f = 3 1,16 1,1 1,2 1,3 2, ,2 3,3 3,16. f = 1. f = Output Layer Using the Gow-And-Pune Netwok to Solve Poblems of Lage Dimensionality B.J. Biedis and T.D. Gedeon School of Compute Science & Engineeing The Univesity of New South Wales Sydney NSW 2052 AUSTRALIA bbiedis@cse.unsw.edu.au

More information

Efficient End-to-End Communication Services for Mixed Criticality Avionics Systems

Efficient End-to-End Communication Services for Mixed Criticality Avionics Systems 214 IEEE 22nd Intenatonal Symposum of Qualty of Sevce (IWQoS) Effcent End-to-End Communcaton Sevces fo Mxed Ctcalty Avoncs Systems Yng Fang Yu Hua Xue Lu McGll Unvesty Huazhong Unv. of Sc. and Tech. McGll

More information

SYSTEM LEVEL REUSE METRICS FOR OBJECT ORIENTED SOFTWARE : AN ALTERNATIVE APPROACH

SYSTEM LEVEL REUSE METRICS FOR OBJECT ORIENTED SOFTWARE : AN ALTERNATIVE APPROACH I J C A 7(), 202 pp. 49-53 SYSTEM LEVEL REUSE METRICS FOR OBJECT ORIENTED SOFTWARE : AN ALTERNATIVE APPROACH Sushil Goel and 2 Rajesh Vema Associate Pofesso, Depatment of Compute Science, Dyal Singh College,

More information

Some Advanced SPC Tools 1. Cumulative Sum Control (Cusum) Chart For the data shown in Table 9-1, the x chart can be generated.

Some Advanced SPC Tools 1. Cumulative Sum Control (Cusum) Chart For the data shown in Table 9-1, the x chart can be generated. Some Advanced SP Tools 1. umulatve Sum ontrol (usum) hart For the data shown n Table 9-1, the x chart can be generated. However, the shft taken place at sample #21 s not apparent. 92 For ths set samples,

More information

QFSRD:OrthogenesisEvolutionBasedGeneticAlgorithmforQosFitnessScopeawareRouteDiscoveryinAd Hoc Networks

QFSRD:OrthogenesisEvolutionBasedGeneticAlgorithmforQosFitnessScopeawareRouteDiscoveryinAd Hoc Networks Global Jounal of Compute Scence and Technology: E Netwo, Web & Secuty Volume 5 Issue 3 Veson.0 Yea 205 Type: Double Blnd Pee Revewed Intenatonal Reseach Jounal Publshe: Global Jounals Inc. (USA) Onlne

More information

Simultaneous Position, Velocity, Attitude, Angular Rates, and Surface Parameter Estimation Using Astrometric and Photometric Observations

Simultaneous Position, Velocity, Attitude, Angular Rates, and Surface Parameter Estimation Using Astrometric and Photometric Observations Smultaneous Poston, Velocty, Atttude, Angula Rates, and Suface Paamete Estmaton Usng Astometc and Photometc Obsevatons Chales J. Wettee and C. Channng Chow Pacfc Defense Solutons, LLC Khe, Hawa, USA jac.wettee@pacfcds.com

More information

Intra-Parametric Analysis of a Fuzzy MOLP

Intra-Parametric Analysis of a Fuzzy MOLP Intra-Parametrc Analyss of a Fuzzy MOLP a MIAO-LING WANG a Department of Industral Engneerng and Management a Mnghsn Insttute of Technology and Hsnchu Tawan, ROC b HSIAO-FAN WANG b Insttute of Industral

More information

Image Enhancement in the Spatial Domain. Spatial Domain

Image Enhancement in the Spatial Domain. Spatial Domain 8-- Spatial Domain Image Enhancement in the Spatial Domain What is spatial domain The space whee all pixels fom an image In spatial domain we can epesent an image by f( whee x and y ae coodinates along

More information

We are IntechOpen, the world s leading publisher of Open Access books Built by scientists, for scientists. International authors and editors

We are IntechOpen, the world s leading publisher of Open Access books Built by scientists, for scientists. International authors and editors We ae IntechOpen, the wold s leadng publshe of Open Access books Bult by scentsts, fo scentsts 3,900 116,000 10 Open access books avalable Intenatonal authos and edtos Downloads Ou authos ae among the

More information

Collision Avoidance for Multiple Robots till Next Waypoints through Collision Free Polygons

Collision Avoidance for Multiple Robots till Next Waypoints through Collision Free Polygons Collson Avodance fo Multple Robots tll Next Wayponts though Collson Fee Polygons Satsh Peddu* K. Madhava Kshna* * Robotcs Reseach Cente, Intenatonal Insttute of Infomaton Technology, Hydeabad, Inda peddu@eseach.t.ac.n

More information

Dual Tree Complex Wavelet Transform for Face Recognition Using PCA Algorithm

Dual Tree Complex Wavelet Transform for Face Recognition Using PCA Algorithm B.Dhamenda,INDIA / Intenatonal Jounal of Reseach and Computatonal echnology, Vol.5 Issue.4 ISSN: 0975-5662, Dec, 2013 www.jct.og Dual ee Complex Wavelet ansfom fo Face Recognton Usng PCA Algothm Dhamenda

More information