Locally Adaptive Dimensionality Reduction for Indexing Large Time Series Databases

Size: px
Start display at page:

Download "Locally Adaptive Dimensionality Reduction for Indexing Large Time Series Databases"

Transcription

1 Locally Adaptve Dmesoalty educto for Idexg Large Tme Seres Databases Kaushk Chakrabart Eamo Keogh Sharad Mehrotra Mchael Pazza Mcrosoft esearch Uv. of Calfora Uv. of Calfora Uv. of Calfora edmod, WA 985 versde, CA 95 Irve, CA 9697 Irve, CA Abstract Smlarty search large tme seres databases has attracted much research terest recetly. It s a dffcult problem because of the typcally hgh dmesoalty of the data.. The most promsg solutos volve performg dmesoalty reducto o the data, the dexg the reduced data wth a multdmesoal dex structure. May dmesoalty reducto techques have bee proposed, cludg Sgular Value Decomposto (SVD), the Dscrete Fourer trasform (DFT), ad the Dscrete Wavelet Trasform (DWT). I ths work we troduce a ew dmesoalty reducto techque whch we call Adaptve Pecewse Costat Approxmato (APCA). Whle prevous techques (e.g., SVD, DFT ad DWT) choose a commo represetato for all the tems the database that mmzes the global recostructo error, APCA approxmates each tme seres by a set of costat value segmets of varyg legths such that ther dvdual recostructo errors are mmal. We show how APCA ca be dexed usg a multdmesoal dex structure. We propose two dstace measures the dexed space that explot the hgh fdelty of APCA for fast searchg: a lower boudg Eucldea dstace approxmato, ad a o-lower boudg, but very tght Eucldea dstace approxmato ad show how they ca support fast exact searchg, ad eve faster approxmate searchg o the same dex structure. We theoretcally ad emprcally compare APCA to all the other techques ad demostrate ts superorty. Categores ad Subject Descrptors: H.3.3 [Iformato Search ad etreval] Search Process H..4 [Systems] Multmeda databases. Keywords: Idexg, dmesoalty reducto, tme-seres smlarty retreval.. Itroducto Tme seres accout for a large proporto of the data stored facal, medcal ad scetfc databases. ecetly there has bee much terest the problem of smlarty search (query-bycotet) tme seres databases. Smlarty search s useful ts ow rght as a tool for exploratory data aalyss, ad t s also a mportat elemet of may data mg applcatos such as clusterg [3], classfcato [6, 33] ad mg of assocato rules []. The smlarty betwee two tme seres s typcally measured wth Eucldea dstace, whch ca be calculated very effcetly. However the volume of data typcally ecoutered exasperates the problem. Mult-ggabyte datasets are very commo. As typcal example, cosder the MACHCO project. Ths database cotas more tha a terabyte of data ad s updated at the rate of several ggabytes a day [48]. CONTACT AUTHO. [Address: Oe Mcrosoft Way, edmod, WA , USA. Phoe: (45) Fax: (45) ] Work doe whle author was Ph.D. studet at Uversty of Illos at Urbaa Champag Work doe whle author was Ph.D. studet at Uversty of Calfora at Irve

2 The most promsg smlarty search methods are techques that perform dmesoalty reducto o the data, the use a multdmesoal dex structure to dex the data the trasformed space. The techque was troduced [] ad exteded [39, 3,]. The orgal work by Agrawal et. al. utlzes the Dscrete Fourer Trasform (DFT) to perform the dmesoalty reducto, but other techques have bee suggested, cludg Sgular Value Decomposto (SVD) [8, 4, 3], the Dscrete Wavelet Trasform (DWT) [9, 49, ] ad Pecewse Aggregate Approxmato (PAA) [4, 5]. For a gve dex structure, the effcecy of dexg depeds oly o the fdelty of the approxmato the reduced dmesoalty space. However, choosg a dmesoalty reducto techque, we caot smply choose a arbtrary compresso algorthm. What s requred s a techque that produces a dexable represetato. For example, may tme seres ca be effcetly compressed by delta ecodg, but ths represetato does ot led tself to dexg. I cotrast SVD, DFT, DWT ad PAA all led themselves aturally to dexg, wth each egewave, fourer coeffcet, wavelet coeffcet or aggregate segmet mappg oto oe dmeso of a dex tree. The ma cotrbuto of ths paper s to propose a smple, but hghly effectve compresso techque, Adaptve Pecewse Costat Approxmato (APCA), ad show that t ca be dexed usg a multdmesoal dex structure. Ths represetato was cosdered by other researchers, but they suggested t does ot allow for dexg due to ts rregularty [5]. We wll show that dexg APCA s possble, ad, usg APCA s up to oe to two orders of magtude more effcet tha alteratve techques o real world datasets. We wll defe the APCA represetato detal Secto 3, however a tutve uderstadg ca be gleaed from Fgure. APCA epresetato ecostructo Error 4.6 DFT ecostructo Error 5.85 Haar Wavelet ecostructo Error 5.77 SVD ecostructo Error Fgure : A vsual comparso of the tme seres represetato proposed ths work (APCA), ad the 3 other represetatos advocated the lterature. For far comparso, all represetatos have the same compresso rato. The recostructo error s the Eucldea dstace betwee the orgal tme seres ad ts approxmato. There are may stuatos whch a user would be wllg to sacrfce some accuracy for sgfcat speedup [5]. Wth ths md we troduce two dstace measures defed o the APCA represetato. The frst tghtly lower bouds the Eucldea dstace metrc ad s used to produce exact earest eghbors. The secod s ot lower boudg, but produces a very close approxmato of Eucldea dstace ad ca be used to quckly fd approxmate earest eghbors. Both methods ca be supported by the same dex structure so that a user ca swtch betwee fast exact search ad eve faster approxmate search. Addtoally we wll show that the APCA represetato ca support queres where the dstace measure s a arbtrary L p orm (.e. p =,,, ). The rest of the paper s orgazed as follows. I Secto we provde backgroud o ad revew related work tme seres smlarty search. I Secto 3 we troduce the APCA represetato, a algorthm to compute t effcetly ad two dstace measures defed o t. I Secto 4 we demostrate how to dex the APCA represetato. Secto 5 cotas a comprehesve expermetal comparso of APCA wth all the competg techques. I secto 6 we dscuss several advatages APCA has over the competg techques, addto to beg faster. Secto 7 offers coclusos ad drectos for future work.

3 . Backgroud ad elated Work Gve two tme seres Q = {q,,q } ad C = {c,,c } ther Eucldea dstace s defed as: ( C ) ( ) Fgure shows the tuto behd the Eucldea dstace. D Q, q c () = C D(Q,C) Q Fgure : The tuto behd the Eucldea dstace. The Eucldea dstace ca be vsualzed as the square root of the sum of the squared legths of the gray les. There are essetally two ways the data mght be orgazed [6]: Whole Matchg.. Here t assumed that all sequeces to be compared are the same legth. Subsequece Matchg. Here we have a query sequece Q (of legth ), ad a loger sequece C (of legth m). The task s to fd the subsequece C of legth, begg at c, whch best matches Q, ad report ts offset wth C. Whole matchg requres comparg the query sequece to each caddate sequece by evaluatg the dstace fucto ad keepg track of the sequece wth the lowest dstace. Subsequece matchg requres that the query Q be placed at every possble offset wth the loger sequece C. Note t s possble to covert subsequece matchg to whole matchg by sldg a wdow of legth across C, ad makg copes of the (m-) wdows. Fgure 3 llustrates the dea. Although ths causes storage redudacy t smplfes the otato ad algorthms so we wll adopt ths polcy for the rest of ths paper. datapots There are two mportat kds of queres that we would lke to support tme seres database, rage queres (e.g., retur all sequeces wth a epslo of the query sequece) ad earest eghbor (e.g., retur the K closest sequeces to the query sequece). The brute force approach to aswerg these queres, sequetal scag, requres comparg every tme seres c to Q. Clearly ths approach s urealstc for large datasets. Ay dexg scheme that does ot exame the etre dataset could potetally suffer from two problems, false alarms ad false dsmssals. False alarms occur whe objects that appear to be close the dex are actually dstat. Because false alarms ca be removed a post-processg stage (by cofrmg dstace estmates o the orgal data), they ca be tolerated so log as they are relatvely frequet. A false dsmssal s whe qualfyg objects are mssed because they appear dstat dex space. We wll refer to smlarty-searchg techques that guaratee o false dsmssals as exact, ad techques that do ot have ths guaratee as approxmate. Approxmate techques ca stll be very useful for explorg large databases, partcularly f the probablty of false dsmssal s low. We wll revew approxmate techques secto. ad exact techques secto.. C C C Fgure 3: The subsequece matchg problem ca be coverted to the whole matchg problem by sldg a "wdow" of legth across the log sequece ad makg copes of the data fallg wth the wdows. 3

4 . Approxmate techques for smlarty searchg. Several researchers have suggested abadog the sstece o exact search favor of a much faster search that returs approxmately the same results. Typcally ths volves trasformg the data wth a lossy compresso scheme, the dog a sequetal search o the compressed data. Typcal examples clude [4, 7, 3, 46], who all utlze a pecewse lear approxmato. Others have suggested trasformg the data to a dscrete alphabet ad usg strg-matchg algorthms [,, 34, 9,, 38]. All these approaches suffer from some lmtatos. They are all evaluated o small datasets resdg ma memory, ad t s uclear f they ca be made to scale to large databases. Further, the systems are evaluated wthout cosderg precso ad recall, thus we ca say lttle or othg about the qualty of the retured aswer set. The work of [3, 36, 45, 5, 6] dffers from the above that they focus provdg a more flexble query laguage ad ot o performace ssues.. Exact techques for smlarty searchg. A tme seres C = {c,,c } wth datapots ca be cosdered as a pot -dmesoal space. Ths mmedately suggests that tme seres could be dexed by multdmesoal dex structure such as the -tree ad ts may varats [7]. Sce realstc queres typcally cota to, datapots (.e. vares from to ) ad most multdmesoal dex structures have poor performace at dmesoaltes greater tha 8- [6], we eed to frst perform dmesoalty reducto order to explot multdmesoal dex structures to dex tme seres data. I [6] the authors troduced GEerc Multmeda INdexIg method (GEMINI) whch ca explot ay dmesoalty reducto method to allow effcet dexg. The techque was orgally troduced for tme seres, but has bee successfully exted to may other types of data [8]. A mportat result [6] s that the authors proved that order to guaratee o false dsmssals, the dstace measure the dex space must satsfy the followg codto: D dex space (A,B) D true (A,B) Ths theorem s kow as the lower boudg lemma or the cotractve property. Gve the lower boudg lemma, ad the ready avalablty of off-the-shelf multdmesoal dex structures, GEMINI requres just the followg three steps. Establsh a dstace metrc from a doma expert ( ths case Eucldea dstace). Produce a dmesoalty reducto techque that reduces the dmesoalty of the data from to N, where N ca be effcetly hadled by your favorte dex structure. Produce a dstace measure defed o the N dmesoal represetato of the data, ad prove that t obeys D dex space (A,B) D true (A,B). Table cotas a outle of the GEMINI dexg algorthm. All sequeces the dataset C are trasformed by some dmesoalty reducto techque ad the dexed by the dex structure of choce. The dexg tree represets the trasformed sequeces as pots N dmesoal space. Each pot cotas a poter to the correspodg orgal sequece o dsk. We remove the mea (optoally) because, for may applcatos, we are oly terested the smlarty based o the shape of the sequece ad ot ts vertcal offset from the x-axs. If the offset s ot removed, t would domate the Eucldea dstace fucto leadg to ututuve otos of smlarty [9]. We remove the mea for the expermets ths paper. For the llustratve examples we use ths paper, we do ot remove the mea for smplcty. 4

5 Algorthm BuldIdex(C,); beg. for all objects database // C s the dataset, s the sze of the wdow. C C Mea(C ); // Optoal: remove the mea of C 3. C SomeTrasformato(C );// C s ay dmesoalty reduced represetato 4. Isert C to the Spatal Access Method wth a poter to C o dsk; 5. edfor ed Table : A outle of the GEMINI dexg buldg algorthm. Note that each sequece has ts mea subtracted before dexg. Ths has the effect of shftg the sequece the y-axs such that ts mea s zero, removg formato about ts offset. Ths step s cluded because for most applcatos the offset s rrelevat whe computg smlarty. Table below cotas a outle of the GEMINI rage query algorthm. Algorthm agequery(q,ε) beg. Project the query Q to the same feature space as the dex.. Fd all caddate objects the dex wth ε of the query. 3. etreve from dsk the actual sequeces poted to by the caddates. 4. Compute the actual dstaces, ad dscard false alarms. ed Table : The GEMINI rage query algorthm. The rage query algorthm s called as a subroute the K Nearest Neghbor algorthm outled Table 3. There are several optmzatos to ths basc K Nearest Neghbor algorthm that we utlze ths paper [4]. We wll dscuss them more detal Secto 4. Algorthm K_NearestNeghbor(Q,K) beg. Project the query Q to the same feature space as the dex.. Fd the K earest caddate objects the dex. 3. etreve from dsk the actual sequeces poted to by the caddates. 4. Compute the actual dstaces ad record the maxmum, call t εmax. 5. Issue the rage query, agequery(q,εmax); 6. Compute the actual dstaces, ad choose the earest K. ed Table 3: The GEMINI earest eghbor algorthm. The effcecy of the GEMINI query algorthms depeds oly o the qualty of the trasformato used to buld the dex. The tghter the boud o D dex space (A,B) D true (A,B) the better, as tghter bouds mply fewer false alarms hece lower query cost [7]. Tme seres are usually good caddates for dmesoalty reducto because they ted to cota hghly correlated features. For brevty, we wll ot descrbe the three ma dmesoalty reducto techques, SVD, DFT ad DWT, detal. Istead we refer the terested reader to the relevat papers or to [4] whch cotas a survey of all the techques. We wll brefly revst related work Secto 6 whe the reader has developed more tuto about our approach. 5

6 3. Adaptve esoluto epresetato I recet work Keogh et. al. [4] ad Y & Faloutsos [5] depedetly suggested approxmatg a tme seres by dvdg t to equal-legth segmets ad recordg the mea value of the datapots that fall wth the segmet. The authors use dfferet ames for ths represetato ([4] calls t Pecewse Aggregate Approxmato whle [5] calls t Segmeted-Meas ), we wll refer to t as Pecewse Aggregate Approxmato (PAA) ths paper. Ths smple techque s surprsgly compettve wth the more sophstcated trasforms. The fact that each segmet PAA s the same legth facltates dexg of ths represetato. Suppose however we relaxed ths requremet ad allowed the segmets to have arbtrary legths, does ths mprove the qualty of the approxmato? Before we cosder ths questo, we must remember that the approach that allows arbtrary legth segmets, whch we call Adaptve Pecewse Costat Approxmato (APCA), requres two umbers per segmet. The frst umber records the mea value of all the datapots segmet, the secod umber records the legth. So a far comparso s N PAA segmets to M APCA segmets, were M = N/. It s dffcult to make ay tutve guess about the relatve performace of the two techques. O oe had PAA has the advatage of havg twce as may approxmatg segmets. O the other had APCA has the advatage of beg able to place a sgle segmet a area of low actvty ad may segmets areas of hgh actvty. I addto oe has to cosder the structure of the data questo. It s possble to costruct artfcal datasets where oe approach has a arbtrarly large recostructo error, whle the other approach has recostructo error of zero. Fgure 4 llustrates a far comparso betwee the two techques o several real datasets. Note that for the task of dexg, subjectve feelgs about whch techque looks better are rrelevat. All that matters s the qualty of the approxmato, whch s gve by the recostructo error (because lower recostructo errors result tghter bouds o D dex space(a,b) D true (A,B).). (A) APCA E = 4.8 (B) APCA E = 98.5 (C) APCA E = 6.9 PAA E = 7. PAA E = 9 PAA E = 63.7 (D) APCA E = 57.3 (E) APCA E = 58. (F) APCA E = 5. PAA E = 7 PAA E = PAA E = 53.4 Fgure 4: A comparso of the recostructo errors of the equal-sze segmet approach (PAA) ad the varable legth segmet approach (APCA), o a collecto of mscellaeous datasets. A) INTEBALL Plasma processes. B) Darw sea level pressures. C) Space Shuttle telemetry. D) Electrocardogram. E) Maufacturg. F) Exchage rate. O fve of the sx tme seres APCA outperforms PAA sgfcatly. Oly o the Exchage ate data are they essetally the same. I fact, we repeated smlar expermets for more tha 4 dfferet tme seres datasets, over a rage of sequece legths ad compresso ratos ad we foud that APCA s always at least as good as PAA, ad usually much better. Ths comparso motvates our approach. If the APCA represetato ca be dexed, ts hgh fdelty to the orgal sgal should allow very effcet prug of the dex space (.e. few false alarms, hece 6

7 low query cost). We wll show how APCA ca be dexed the ext secto (Secto 4). I the rest of ths secto, we defe the APCA represetato formally, descrbe the algorthm to obta the APCA represetato of a tme seres ad dscuss the dstace measures for APCA. 3. The APCA represetato Gve a tme seres C = {c,,c }, we eed to be able to produce a APCA represetato, whch we wll represet as C ={<cv,cr >,,<cv M,cr M >}, cr = () cr cr 3 cr cv 3 cr 4 C cv cv C cv 4 Fgure 5: A tme seres C ad ts APCA represetato C, wth M = 4 where cv s the mea value of datapots the th segmet (.e. cv = mea( c cr c +,..., cr )) ad cr the rght edpot of the th segmet. We do ot represet the legth of the segmets but record the locatos of ther rght edpots stead for dexg reasos as wll be dscussed Secto 4. The legth of the th segmet ca be calculated as cr cr -. Fgure 5 llustrates ths otato. Symbols Deftos S Number of objects the database. Legth of tme seres (a.k.a. query legth, orgal dmesoalty) C = {c,,c } A tme seres a database, stored a vector of legth. Q = {q,,q } A query tme seres, represeted as a vector of legth. N Dmesoalty of dex structure, wth N <<. M Number of segmets APCA represetato, wth M = N/. C = A adaptve pecewse costat approxmato of C, wth c the mea value { <cv,cr >,,<cv M,cr M > } of th segmet ad cr the rght edpot of th segmet. Q = Also a adaptve pecewse costat approxmato, but obtaed usg a { <qv,qr >,,<qv M,qr M > } specal algorthm as descrbe Equato 4 D(Q,C) Eucldea dstace D AE (Q,C) A o-lower boudg approxmato of Eucldea dstace D LB (Q,C) or D LB (Q,C) A lower boudg approxmato of the Eucldea dstace cmax, cm The max ad m values of APCA represetato C th segmet =(L, H)= MB assocated wth a ode (say U) of the dex bult o N-dmesoal ({l,,l N }, {h,,h N }) APCA space; L={ l,,l N } ad H={h,,h N } deote the lower ad hgher edpots of the major dagoal of. C = ({cm, cr,, cm M, cr M }, { APCA rectagle correspodg to APCA pot C cmax, cr,, cmax M, cr M }) G = th rego assocated wth ; G [] ad G [3] are low ad hgh bouds {G [], G [], G [3], G [4]} alog the value axs; G [] ad G [4] are those alog the tme axs MINDIST(Q, ) Mmum dstace of MB from query tme seres Q MINDIST(Q,, t) Mmum dstace of MB from Q at tme stat t MINDIST(Q, G, t) Mmum dstace of rego G from Q at tme stat t Table 4: The otato used ths paper. 3. Obtag the APCA represetato As metoed before, the performace of the dex structure bult o the APCA represetato defed Equato depeds o how closely the APCA represetato approxmates the orgal sgal. Closer the approxmato, fewer the umber of false alarms, better the performace of the dex. We say that a M-segmet APCA represetato C of a tme seres C s optmal ( terms of the qualty of approxmato) ff C has the least recostructo error amog all possble M- segmet APCA represetatos of C. Fdg the optmal pecewse polyomal represetato of 7

8 a tme seres requres a O(M ) dyamc programmg algorthm [5, 35]. Ths s too slow for hgh dmesoal data. I ths paper, we propose a ew algorthm to produce almost optmal APCA represetatos O(log()) tme. The algorthm works by frst covertg the problem to a wavelet compresso problem, for whch there are well kow optmal solutos, the covertg the soluto back to the ACPA represetato ad (possbly) makg mor modfcatos. The algorthm leverages off the fact that the Haar wavelet trasformato of a tme seres sgal ca be calculated O(), ad that a optmal recostructo (.e., havg least recostructo error) of the sgal for ay level of compresso (.e., #retaed coeffcets/) ca be obtaed by sortg the coeffcets order of decreasg ormalzed magtude, the trucatg off the smaller coeffcets [44]. The segmets the recostructed sgal may have approxmate mea values (due to trucato); we replace them by the exact mea values to get a vald APCA represetato as defed Equato. There are, however, two ssues we must address before utlzg ths approach. ) The DWT s defed oly for tme seres wth a legth that s a teger power of two whle may ot ecessarly be a power of two. Ths problem ca be solved easly by paddg those tme seres wth zeros, the trucatg the correspodg segmet after performg the DWT. ) There s o exact mappg betwee the umber of Haar coeffcets retaed ad the umber of segmets the APCA represetato resultg from the recostructo. For example a sgle coeffcet Haar approxmato could produce a, or 3-segmet APCA represetato. Our soluto s to keep the largest M coeffcets; ths wll produce a APCA represetato wth the umber of segmets betwee M ad 3M. If the umber of segmets s more tha M, adjacet pars of segmets are merged utl exactly M segmets rema. The segmet pars targeted for mergg are those that ca be fused to a sgle segmet wth the mmum crease recostructo error. Table 5 cotas the outle of the algorthm. Algorthm Compute_APCA(C,M) beg. f legth(c) s ot a power of two, pad t wth zeros to make t so.. Perform the Haar Dscrete Wavelet Trasform o C. 3. Sort coeffcets order of decreasg ormalzed magtude, trucate after M. 4. ecostruct approxmato (APCA represetato) of C from retaed coeffs. 5. If C was padded wth zeros, trucate t to the orgal legth. 6. eplace approxmate segmet mea values wth exact mea values. 7. whle the umber of segmets s greater tha M 8. Merge the par of segmets that ca be merged wth least rse error 9. edwhle ed Table 5: A algorthm to produce the APCA. The parameter M should be chose judcously. If M s too large, the dmesoalty N of dex structure (N=M) wll be hgh resultg hgh query cost (due to dmesoalty curse). If M s too small, the recostructo error may become large leadg to too may false postves ad hece hgh query cost. So, M should be chose such that the overall recostructo error remas low wthout lettg the dmesoalty exceed the crtcal threshold of the dex structure (above whch t performs worse tha sequetal sca). The actual choce of M would deped o the dataset ad the multdmesoal dex structure used. 8

9 We llustrate the workg of the above algorthm usg a umercal example. Example (Computg APCA represetato) Let us cosder a tme seres C=[7, 5, 5, 3, 3, 3, 4, 6]. Table 6 shows the Haar wavelet decomposto of the seres. We start by parwse averagg the values to get a ew lowerresoluto represetato of the data wth the followg average values [(7+5)/, (5+3)/, (3+3)/, (4+6)/] = [6, 4, 3, 5]. Obvously, some formato s lost ths averagg process. To be able to recostruct the orgal seres, we eed to also store the dffereces of the (secod of the) averaged values from the computed parwse average,.e., [6-5, 4-3, 3-3, 5-6] = [,,, -]. These are called the detal coeffcets. Applyg the parwse averagg ad dfferecg process recursvely o the lower-resoluto array cotag the averages, we get the full decomposto show Table 6. esoluto Averages Detal Coeffcets 3 [7, 5, 5, 3, 3, 3, 4, 6] - [6, 4, 3, 5] [,,, -] [5, 4] [, -] [4.5] [.5] Table 6: Haar Wavelet Trasform for APCA Computato The wavelet trasform W C of C cossts of the sgle coeffcet represetg the overall average of data values followed by the detaled coeffcets the order of creasg resoluto. So W C = [4.5,.5,, -,,,, -]. To take to accout the mportace of a coeffcet wth regard to the recostructo of the orgal seres (.e., the umber of elemets the seres t cotrbutes to the recostructo of), we ormalze the trasform by dvdg each coeffcet by (l/) where l s the level of resoluto of the coeffcet. So the ormalzed wavelet trasform of C s W = [ 4. 5,. 5,,,,,, ]. 7 C Segm et 5.5 Segm et Segm et 3.5 Segm et (a ) (b ) Segm et cv = 6 Segm et cv = 4 Segm et 3 cv 3 = 3 Segm et 4 cv 4 = 5 Segm et cv = 6 Segm et cv = 3.5 Segm et 3 cv 3 = 5 (c ) (d ) Fgure 6: Step-by-step workg of Compute_APCA algorthm. (a) Orgal tme seres C = [7,5,5,3,3,3,4,6] (b) Tme seres recostructed from the M (3 ths case) best wavelet coeffcets of C. The recostructed seres has 4 segmets (segmet boudares dcated by dots). The mea value of each segmet s show just above the segmet. (c) ecostructed tme seres wth approxmate meas replaced by exact meas (cv s). (d) Fal APCA represetato obtaed by mergg segmets ad 3 (to reduce the umber of segmets to M=3). 9

10 Suppose M=3. So we would reta the 3 coeffcets wth hghest ormalzed magtude,.e., the frst, thrd ad fourth coeffcets. Fgure 6(a) ad (b) show the orgal tme seres C ad tme seres recostructed from those 3 coeffcets respectvely. Fgure 6(c) shows the recostructed tme seres wth approxmate segmet mea values replaced by the exact oes. Fally, we eed to merge oe par of segmets to reduce the umber of segmets to M=3; segmet ad 3 s the best par to merge as t results the mmum crease recostructo error. Fgure 6(d) shows the fal 3-segmet APCA represetato of C produced by the Compute_APCA algorthm. We expermetally compared ths algorthm wth several of the heurstc, mergg algorthms [5, 35, 4] ad foud t s faster (at least 5 tmes faster for ay legth tme seres) ad slghtly superor terms of recostructo error. 3.3 Dstace measures defed for the APCA represetato Suppose we have a tme seres C, whch we covert to the APCA represetato C, ad a query tme seres Q. Clearly, o dstace measure defed betwee Q ad C ca be exactly equvalet to the Eucldea dstace D(Q,C) (defed Equato ) because C geerally cotas less formato tha C. However, we wll defe two dstace measures betwee Q ad C that approxmate D(Q,C). The frst, D AE (Q,C) s desged to be a very tght approxmato of the Eucldea dstace, but may ot always lower boud the Eucldea dstace D(Q,C). The secod, D LB (Q,C) s geerally a less tght approxmato of the Eucldea dstace, but s guarateed to lower-boud, a property ecessary to utlze the GEMINI framework. These dstace measures are defed below, Fgure 7 llustrates the tuto behd the formulas A approxmate Eucldea measure D AE Gve a query Q, raw data format, ad a tme seres C the APCA represetato, D AE (Q,C) s defed as: M cr cr cv q (3) D AE (Q,C) ( k cr ) = k= + Ths measure ca be effcetly calculated O(), ad t tghtly approxmates the Eucldea dstace, ufortuately t has a drawback whch prevets ts use for exact search. Proposto D AE (Q,C) does ot satsfy the tragular equalty Proof: By couter example. The tragular equalty states that for ay objects α, β ad χ d(α,β) d(α,χ) + d(β,χ) I other words, f D AE (Q,C) obeys tragle equalty, there ca exst o object A, B ad C such that D AE (A,B) D AE (A,C) + D AE (B,C) We prove our proposto by fdg three such objects. Cosder the tme seres A = {-, -, -,, }, B = {,,, -, -} ad C = {,,,, -}. The - segmet APCA represetatos of A, B ad C as produced by the Compute_APCA algorthm are A = {<-,>, < / 3,5>}, B = {< / 3,3>,<-,5>} ad C = {< /,>,< - / 3,5>} respectvely. Accordg to Equato 3, D AE (A,B) (/3- (-)) + (/3- (-)) + (/3- (-)) + (--) + (-- ) Smlarly, D AE (A,C) = ad D AE (B,C) =.47. So, D AE (A,C) + D AE (B,C) = D AE (A,B) D AE (A,C) + D AE (B,C) Ths mples D AE (Q,C) does ot satsfy the tragular equalty.

11 The falure of D AE to obey the tragular equalty meas that t may ot lower boud the Eucldea dstace ad thus caot be used for exact dexg [5]. However, we wll demostrate later that t s very useful for approxmate search A lower-boudg measure D LB To defe D LB (Q,C) we must frst troduce a specal verso of the APCA. Normally the algorthm metoed Secto 3. s used to obta ths represetato. However we ca also obta ths represetato by projectg the edpots of C oto Q, ad fdg the mea value of the sectos of Q that fall wth the projected tervals. A tme seres Q coverted to the APCA represetato ths way s deoted as Q. The dea ca be vsualzed Fgure 7 III. Q s defed as: Q ={<qv,qr >,,<qv M,qr M >}, where qr = cr ad qv = mea( q cr q +,..., cr ) (4) M D LB (Q,C) s defed as: D LB (Q,C) = ( cr cr )( qv cv (5) ) I Q III C C Q Q II IIII D AE ( Q,C ) D LB (Q,C ) Fgure 7: A vsualzato of the two dstace measures defe o the APCA represetato. I) A query tme seres Q ad a APCA object C. II) The D AE measure ca be vsualzed as the Eucldea dstace betwee Q ad the recostructo of C. III) Q s obtaed by projectg the edpots of C oto Q ad calculatg the mea values of the sectos fallg wth the projected les. IIII) The D LB measure ca be vsualzed as the square root of the sum of the product of squared legth of the gray les wth the legth of the segmets they jo. We llustrate the computato of D LB (Q,C) usg a umercal example below. Example (Computato of D LB (Q,C)) Let us cosder a tme seres A={4, 6,,, }. The -segmet APCA represetato of A as produced by the Compute_APCA algorthm s A={<5,>, <,5>}. Let Q = {5, 3, 5, 6, 7} be a query tme seres. To compute D LB (Q,C), we frst compute Q ={<4,>, <6,5>}. D LB (Q,C) = ( )(4 5) + (5 )(6 ) = Note that D LB (Q,C) lower bouds D(Q,C) = = ( 5 4) + (3 6) + (5 ) + (6 ) + (7 ) = The formal proof s show below. Lemma : D LB (Q,C) lower bouds the Eucldea Dstace D(Q,C). Proof: We preset a proof for the case where there s a sgle segmet the APCA represetato. The more geeral proof for the M segmet case ca be obtaed by applyg the proof to each of the M segmets. Let W={w, w,, w p } be a vector of p real umbers. Let W deote the arthmetc mea of W,.e., W = (Σ w )/p. We defe a vector W of real umbers where w W w. It s easy to see that Σ w =. The defto of w allows us to substtute w by W w, a fact whch we wll utlze the proof below. Let Q ad C be the query ad data tme seres respectvely, wth Q = C =. Let Q ad C be the correspodg APCA vectors as defed Equatos 4 ad respectvely.

12 We wat to prove M ( q ) = = c ( cr cr qv cv )( ) Because we are cosderg just the sgle segmet case, we ca remove summato over M segmets ad rewrte the equalty as: Assume Because (cr cr - ) = = ( q c ) ( cr cr )( qv cv ) ( ) ( ) = q c qv cv ( ) ( ) Because the terms uder the radcals must be oegatve, we ca square both sdes qv s smply the mea of Q, so rewrte as Q cv s smply the mea of C, so rewrte as C Substtute rearragemet of defto above = = q c qv cv ( q c ) ( Q C ) (( Q q ) ( C c )) ( Q C ) = ( ) ( ) earrage terms ( Q C) ( q c ) Q C = ( ) Bomal theorem Dstrbutve law Summato propertes Assocatve law Σ w =, proved above ( Q C) ( Q C)( q c ) ( q c ) = + Q C ( Q C) ( Q C)( q c ) ( q c ) + = = = ( Q C ) ( Q ) ( Q C) ( Q C) ( q c ) ( q c ) C + = = ( q c ) ( q c ) ( Q ) + = ( Q C) ( Q C) C = = C) ( Q C) ( ) + ( q c ) ( Q ) = ( ) ( Q C Subtract lke term from both sdes ( Q C) + ( q c ) Q C = = The sum of squares must be oegatve, so our assumpto was true. Hece the proof. ( q c ) 3.4 Qualty of proposed dstace measures D LB ad D AE The qualty of a lower boudg dstace measure ca be gauged by how tghtly t lower bouds the true dstace betwee all queres of terest ad all objects the database (because all queres of terest caot be kow advace ad the database may be very large, a large radom samplg must suffce). For a o-lower boudg measure, the qualty s a lttle more dffcult to defe. Itutvely we wat the measure to tghtly approxmate the true dstace but oly rarely overestmate t. I addto, whe the true dstace s overestmated, t should ot be by a large amout.

13 We devsed a smple expermet to llustrate the qualty of D LB ad D AE compared to the DWT (Haar) ad the DFT approaches. We radomly extracted two sequeces A ad B from a database ad measured the true Eucldea dstace D(A,B) betwee them. We also measured the dstace betwee A ad B usg the varous reduced dmesoalty represetatos for a fxed value of N. The rato of the estmated dstace over the true dstace for all combatos was used to plot a pot -space, as llustrated Fgure 8. Sample of Electrocardogram data Sample of Star Lght Curve data ato of estmated dstace over true Eucldea dstace for competg approaches (Haar wavelet ad DFT) ego whch APCA has less prug power ego whch APCA has more prug power ego whch APCA has less prug power ego whch APCA has more prug power ego whch APCA volates the lower boudg lemma ato of estmated dstace over true Eucldea dstace for D LB ato of estmated dstace over true Eucldea dstace for D AE Haar vs APCA DFT vs APCA Electrocardogram Star Lght Curve Fgure 8: A vsualzato of the prug power of the two dstace measures defed o APCA as compared to the prug power of the Haar wavelet ad DFT approaches. Pots o the dagoal dcate that the two approaches beg compared have the same prug power, pots below the dagoal dcate that APCA s superor ad pots above the dagoal dcate that APCAs rval s superor. We repeated ths, tmes wth radomly chose sequeces for each of two datasets. DWT ad DFT behaved smlarly, so for brevty we wll oly dscuss the comparsos betwee DWT ad the two measures defed o APCA. For the Electrocardogram dataset D LB produces tghter lower bouds tha the Haar wavelet approach 99.9% of the tme, ad the dfferece s usually qute sgfcat. The D AE measure very tghtly approxmated the true dstace, ad oly volated lower boudg.9% of the tme, geerally by a very small amout. For the Star Lght Curve dataset D LB produces tghter lower bouds tha the Haar wavelet approach 8.3% of the tme. The D AE measure oly volates lower boudg.% of the tme ad geerally s a extremely tght approxmato of the true dstace. The qualty of these results strogly suggests that APCA would be superor to exstg approaches f dexable. We wll address ths ssue the ext secto. 3

14 4. Idexg the APCA represetato The APCA represetato proposed Secto 3. defes a N-dmesoal feature space (N = M). I other words, the proposed represetato maps each tme seres C = {c,,c } to a pot C = {cv, cr,, cv M, cr M } a N-dmesoal space. We refer to the N-dmesoal space as the APCA space ad the pots the APCA space as APCA pots. I ths secto, we dscuss how we ca dex the APCA pots usg a multdmesoal dex structure (e.g., -tree) ad use the dex to aswer rage ad K earest eghbors (K-NN) queres effcetly. We wll cocetrate o K-NN queres ths secto; rage queres wll be dscussed brefly at the ed of the secto. Algorthm ExactKNNSearch(Q,K) Varable queue: MPrortyQueue; Varable lst: temp; beg. queue.push(root_ode_of_dex, );. whle ot queue.isempty() do 3. top = queue.top(); 4. for each tme seres C temp such that D(Q,C) top.dst 5. emove C from temp; Add C to result; f result = K retur result; 8. edfor 9.. queue.pop(); f top s a APCA pot C. etreve full tme seres C from database;. 3. temp.sert(c, D(Q,C)); else f top s a leaf ode 4. for each data tem C top queue.push(c, D LB (Q,C)); edfor 7. else // top s a o-leaf ode 8. for each chld ode U top 9. queue.push(u, MINDIST(Q,)) // s MB assocated wth U. edfor. edf. eddo ed Table 6: K-NN algorthm to compute the exact K earest eghbors of a query tme seres Q usg a multdmesoal dex structure A K-NN query (Q, K) wth query tme seres Q ad desred umber of eghbors K retreves a set C of K tme seres such that for ay two tme seres C C, E C, D(Q, C) D(Q, E). The algorthm for aswerg K-NN queres usg a multdmesoal dex structure s show Table 6 3. The above algorthm s a optmzato o the GEMINI K-NN algorthm descrbed Table 3 ad was proposed [4]. Lke the basc K-NN algorthm [9,4], the algorthm uses a prorty queue queue to avgate odes/objects the dex the creasg order of ther dstaces from Q the dexed (.e. APCA) space. The dstace of a object (.e. APCA pot) C from Q s defed by D LB (Q,C) (cf. Secto 3.3.) whle the dstace of a ode U from Q s defed by the mmum dstace MINDIST(Q,) of the mmum boudg rectagle (MB) assocated wth U from Q (defto of MINDIST wll be dscussed later). Itally, we push the root ode of the dex to the queue (Le ). Subsequetly, the algorthm avgates the dex by 3 I ths paper, we restrct our dscusso to oly feature-based dex structures.e. multdmesoal dex structures that recursvely cluster pots usg mmum boudg rectagles (MBs). Examples of such dex structures are - tree, X-tree ad Hybrd Tree. Note that the MB-based clusterg ca be logcal.e. the dex structure eed ot store the MBs physcally as log as they ca be derved from the physcally stored formato. For example, space parttog dex structures lke the hb-tree ad the Hybrd Tree store the parttog formato sde the dex odes as kd-trees [4, 6]. Sce the MBs ca be derved from the kd-trees, the techques dscussed here are applcable to such dex structures [6]. 4

15 poppg out the tem from the top of queue at each step (Le 9). If the popped tem s a APCA pot C, we retreve the orgal tme seres C from the database, compute ts exact dstace D(Q,C) from the query ad sert t to a temporary lst temp (Les -). If the popped tem s a ode of the dex structure, we compute the dstace of each of ts chldre from Q ad push them to queue (Les 3-). We move a tme seres C from temp to result oly whe we are sure that t s amog the K earest eghbors of Q.e. there exsts o object E result such that D(Q,E) < D(Q,C) ad result < K. The secod codto s esured by the ext codto Le 7. The frst codto ca be guarateed as follows. Let I be the set of APCA pots retreved so far usg the dex (.e. I = temp result). If we ca guaratee that C I, E I, D LB (Q,C) D(Q,E), the the codto D(Q,C) top.dst Le 4 would esure that there exsts o uexplored tme seres E such that D(Q, E) < D(Q,C). By sertg the tme seres temp (.e. already explored objects) to result creasg order of ther dstaces D(Q,C) (by keepg temp sorted by D(Q,C)), we ca esure that there exsts o explored object E such that D(Q, E) < D(Q,C). I other words, f C I, E I, D LB (Q,C) D(Q,E), the above algorthm would retur the correct aswer. Before we ca use the above algorthm, we eed to descrbe how to compute MINDIST(Q,) such that the correctess requremet s satsfed.e. C I, E I, D LB (Q,C) D(Q,E). We ow dscuss how the MBs are computed ad how to compute MINDIST(Q,) based o the MBs. We start by revstg the tradtoal defto of a MB [7]. Let us assume we have bult a dex of the APCA pots by smply sertg the APCA pots C = {cv, cr,, cv M, cr M } to a MB-based multdmesoal dex structure (usg the sert fucto of the dex structure). Let U be a leaf ode of the above dex. Let = (L, H) be the MB assocated wth U where L = {l, l,, l N } ad H = {h, h,, h N } are the lower ad hgher edpots of the major dagoal of. By defto, s the smallest rectagle that spatally cotas each APCA pot C = {cv, cr,, cv M, cr M } stored U. Formally, = (L, H) s defed as: Defto 4. (Old defto of MB) l = m C U cv( + ) / f s odd (6) L = m cr f s eve C U / h f s odd L = max C U cv( + ) / L = max cr f s eve C U The MB assocated wth a o-leaf ode would be the smallest rectagle that spatally cotas the MBs assocated wth ts mmedate chldre [7]. / cmax3 cmax cmax cm3 cm C cm C Fgure 9: Defto of cm ad cmax for computg MBs cmax4 cm4 However, f we buld the dex as above (.e. the MBs are computed as Defto 4.), t s ot possble to defe a MINDIST(Q,) that satsfes the correctess crtera. To overcome the problem, we defe the MBs are follows. Let us cosder the MB of a leaf ode U. For ay APCA pot C = {cv, cr,, cv M cr M } stored ode U, let cm ad cmax deote the mmum ad maxmum values of the correspodg tme seres C amog the datapots the th segmet.e. 5

16 cr cm = m ( c ) t = cr t + ad (7) cr cmax = max ( c ) t = cr t + The cm ad cmax of a smple tme seres wth 4 segmets s show Fgure 9. We defe the MB = (L, H) assocated wth U as follows: Defto 4. (New defto of MB) l = m C U cm( + ) / f s odd (8) = m cr f s eve C U / h f s odd = max C U cmax( + ) / = max cr f s eve C U As before, the MB assocated wth a o-leaf ode s defed as the smallest rectagle that spatally cotas the MBs assocated wth ts mmedate chldre. How do we buld the dex such that the MBs satsfy Defto 4.. We sert rectagles stead of the APCA pots. I order to sert a APCA pot C = {cv,cr,..,cv M,cr M }, we sert a rectagle C = ({cm, cr,, cm M, cr M },{ cmax, cr,, cmax M, cr M }) (.e. {cm, cr,, cm M, cr M } ad { cmax, cr,, cmax M, cr M }) are the lower ad hgher edpots of the major dagoal of C ) to the multdmesoal dex structure (usg the sert fucto of the dex structure). Sce the serto algorthm esures that the MB of a leaf ode U spatally cotas all the C s stored U, satsfes defto 4. (as llustrated example 3 below). The same s true for MBs assocated wth o-leaf odes. Example 3 (Computato of MBs) Let us cosder two tme seres A={4, 6,,, } ad B={4, 3, 5,, 3}. The -segmet APCA represetatos of A ad B as produced by the Compute_APCA algorthm are A= {<av, ar >, <av, ar >} = {<5,>, <,5>} ad B = {<bv, br >, <bv, br >} ={<4,3>, <,5>} respectvely. For A, am = m(4,6) = 4, amax = max(4,6) = 6, am = m(,,) =, amax = max(,,) =. For B, bm = m(4,3,5) = 3, bmax = max(4,3,5) = 5, bm = m(,3) =, bmax = max(,3) = 3. So, APCA rectagles A = ({am, ar, am, ar }, {amax, ar, amax, ar }) = ({4,,, 5}, {6,,, 5}) ad B = ({bm, br, bm, br }, {bmax, br, bmax, br }) = ({3, 3,, 5}, {5, 3, 3, 5}). Sce the MB of A ad B s the smallest rectagle that spatally cotas A ad B, = ({m(am, bm ), m(ar, br ), m(am, bm ), m(ar, br )}, {max(amax, bmax ), max(ar, br ), max(amax, bmax ), max(ar, br )}) whch satsfes defto 4.. To complete the example, = ({m(4,3), m(,3), m(,), m(5,5)},{max(6,5), max(,3), max(,3), max(5,5)}) = ({3,,, 5}, {6, 3, 3, 5}). Sce we use oe of the exstg multdmesoal dex structures to buld the APCA dex, the storage orgazato of the odes follows that of the dex structure (e.g., MB, chld_ptr array f -tree s used, kd-tree f hybrd tree s used). For the leaf odes, we eed to store the cv s of each data pot ( addto to the cmax s, cm s ad cr s) sce they are eeded to compute D LB (Le 5 of the K-NN algorthm Table 6). The dex ca be optmzed ( terms of leaf ode faout) by ot storg the cmax s ad cm s of the data pots at the leaf odes.e. just storg the cv s ad cr s (a total of M umbers) per data pot addto to the tuple detfer. The reaso s that the cmax s ad cm s are ot requred for computg D LB, ad hece are ot used by the K-NN algorthm. They are eeded just to compute the MBs properly (accordg to defto 4.) at the tme of serto. The oly tme they are eeded later (after the tme of serto) s durg the recomputato of the MB of the leaf ode cotag the data pot after a ode splt. The sert fucto of the dex structure ca be easly modfed to fetch the cmax s / 6

17 ad cm s of the ecessary data pots from the database (usg the tuple detfers) o such occasos. The small extra cost of such fetches durg ode splts s worth the mprovemet search performace due to hgher leaf ode faout. We have appled ths optmzato the dex structure for our expermets but we beleve the APCA dex would work well eve wthout ths optmzato. Oce we have bult the dex as above (.e. the MBs satsfy Defto 4.), we defe the mmum dstace MINDIST(Q,) of the MB assocated wth a ode U of the dex structure from the query tme seres Q. For correctess, C I, E I, D LB (Q,C) D(Q,E) (where I deotes the set of APCA pots retreved usg the dex at ay stage of the algorthm). We show that the above correctess crtera s satsfed f MINDIST(Q,) lower bouds the Eucldea dstace D(Q,C) of Q from ay tme seres C placed uder U the dex. Lemma : If MINDIST(Q,) D(Q,C) for ay tme seres C placed uder U, the algorthm Table 6 s correct.e. C I, E I, D LB (Q,C) D(Q,E) where I deotes the set of APCA pots retreved usg the dex at ay stage of the algorthm. Proof: Accordg to the K-NN algorthm, ay tem E I must satsfy oe of the followg codtos: ) E has bee serted to the queue but has ot bee popped yet.e. C I, D LB (Q, C) D LB (Q,E) ) E has ot yet bee serted to the queue.e. there exsts a paret ode U of E whose MB satsfes the followg codto: C I, D LB (Q,C) MINDIST(Q,). Sce D LB (Q,E) D(Q,E) (Lemma ), () mples C I, D LB (Q,C) D(Q,E). If MINDIST(Q,) D(Q,E) for ay tme seres E uder U, () mples that C I, D LB (Q, C) D(Q,E). Sce ether () or () must be true for ay tem E I, C I, E I, D LB (Q,C) D(Q,E). A trval defto MINDIST(Q,) that lower bouds D(Q,C) for ay tme seres C uder U s MINDIST(Q,) = for all Q ad. However, ths defto s too coservatve ad would cause the K-NN algorthm to vst all odes of the dex structure before returg ay aswer (thus defeatg the purpose of dexg). The larger the MINDIST, the more the umber of odes the K-NN algorthm ca prue, the better the performace. We provde such a defto of MINDIST below 3. Let us cosder a ode U wth MB = (L,H). We ca vew the MB as two APCA represetatos L={<l, l >,, <l N-, l N >} ad H = {<h, h >,, < h N-, h N >}. The vew of a 6- dmesoal MB ({l,l,..,l 6 }, {h,h,,h 6 }) as two APCA represetatos {<l, l >, < l 5, l 6 >} ad {<h, h >, <h 5, h 6 >} s show Fgure. Ay tme seres C = {c, c,, c } uder the ode U s cotaed wth the two boudg tme seres L ad H (as show Fgure ). I order to formalze ths oto of cotamet, we defe a set of M regos assocated wth. The th rego G ( =,, M) assocated wth s defed as the -dmesoal rectagular rego the value-tme space that fully cotas the th segmet of all tme seres stored uder U. The boudary of a rego G, beg Note that MINDIST (Q,) does ot have to lower boud D LB (Q,C) for ay C uder U; t just has to lower boud D(Q,C) for ay C uder U. 3 Idex structures ca allow exteral applcatos to plug doma-specfc MINDIST fuctos ad pot-to-pot dstace fuctos ad retreve earest eghbors based o those fuctos (e.g., Cosstet fucto GST). 7

18 EGION G = {l 3, l +, h 3, h 4 } t t H= { <h, h >, <h 3, h 4 >, <h 5, h 6 >} value axs h l 3 h 3 l h 5 EGION 3 G 3 = {l 5, l 4 +, h 5, h 6 } l 5 Ay tme seres C={c,..., c } uder ths ode wth MB=(L,H) EGION l G = {l,, h, h } l 4 h h l s cotaed betwee L ad H 4 6 (the dots o the tme seres mark h 6 the starts ad eds of the 3 tme axs L= { <l, l >, <l 3, l 4 >, <l 5, l 6 >} APCA segmets Fgure : The M egos assocated wth a M-dmesoal MB. The boudary of a rego G s deoted by G = {G[], G[], G[3], G[4]} a -d rectagle, s defed by 4 umbers: the low bouds G[] ad G[] ad the hgh bouds G[3] ad G[4] alog the value ad tme axes respectvely. By defto, G ] = m ( cm ) (9) [ C uder U G [ ] m ( cr + ) = C uder U G [ 3] = maxc uder U ( cmax ) G 4] = max ( cr ) [ C uder U Based the defto of MB Defto 4., G ca be defed terms of the MB as follows: Defto 4.3 (Defto of regos assocated wth MB) G [] = l {-} () G [] = l {-} + G [3] = h {-} G [4] = h {} Fgure shows the 3 regos assocated wth the 6-dmesoal MB ({l,l,..,l 6 }, {h,h,,h 6 }). We llustrate the rego computato usg a umerc example Example 4. Example 4 (ego Computato) Let us cosder the MB Example 3. ecall =({l, l, l 3, l 4 }, {h, h, h 3, h 4 }) = ({3,,, 5),{6, 3, 3, 5}). The two regos assocated wth are: G = {m(am, bm ), m(ar +, br +), max(amax, bmax ), max(ar, br )} (by Equato 9) = {l, l +, h, h } (by Equato ) = {3,, 6, 3} G = {m(am, bm ), m(ar +, br +), max(amax, bmax ), max(ar, br )} ={l 3, l +, h 3, h 4 } = {,, 3, 5} 8

19 At tme stace t (t =,,), we say a rego G s actve ff G [] t G [4]. For example, Fgure, oly regos ad are actve at tme stat t whle regos, ad 3 are actve at tme stat t. The value c t of a tme seres C uder U at tme stat t must le wth oe of the regos actve at t.e. G [] c G s actve t G [3]. Lemma 3:The value c t of a tme seres C uder U at tme stat t must le wth oe of the regos actve at t. Proof: Let us cosder a rego < t. Frst, let us cosder the case uder U. Sce G that s ot actve at tme stat t.e. ether G [] > t. By defto, G [] > t or G [4] G [] cr - + for ay C G [] > t, t < cr - +.e. c t s ot segmet. Now let us cosder the case G [4] < t. By defto, G [4] cr for ay C uder U. Sce G [4] < t, t > cr.e. c t s ot segmet. Hece, f rego G s ot actve at t, c t caot le segmet.e. c t ca le segmet oly f G s actve. By defto of regos, c t must le wth oe of the regos actve at t.e. G s actve G [] c t G [3]. EGION G = {l 3, l +, h 3, h 4 } value axs Q uery tm e seres Q = {q,..., q } EGION 3 G 3 = {l 5, l 4 +, h 5, h 6 } EGION G = {l,, h, h } tm e axs MINDIST(Q,,t) =m (M IN D IST(Q,G,t), M IN D IST (Q,G,t)) =m ((q t - h), (q t - h3) ) =(q t - h3) t Fgure : Computato of MINDIST Gve a query tme seres Q = {q, q,, q }, the mmum dstace MINDIST(Q,,t) of Q from at tme stat t (cf. Fgure ) s gve by m MINDIST(Q,G,t) where rego G s actve at t MINDIST(Q,G,t) = (G[]-q t ) f q t < G[] () = (q t -G[3]) f G[3] < q t = otherwse. MINDIST(Q,) s defed as follows: MINDIST(Q,) = t = We llustrate the MINDIST computato usg a umerc example Example 5. Example 5 (MINDIST Computato) Let us cosder the MB Examples 3 ad 4 ad ts assocated regos cosder the query tme seres Q = {5, 3, 5, 6, 7} Example. MINDIST(Q,,) = MINDIST(Q, G, ) = t MINDIST(Q,,t) =m (M IN D IST(Q,G,t), M IN D IS T (Q,G,t), M IN D IST (Q,G 3,t )) =m ((q t - h),, (q t - h3) ) = MINDIST( Q,, t) () G ad G. Let us 9

20 MINDIST(Q,,) = m(mindist(q, MINDIST(Q,,3) = m(mindist(q, MINDIST(Q,,4) = MINDIST(Q, MINDIST(Q,,5) = MINDIST(Q, G, ), MINDIST(Q, G, 3), MINDIST(Q, G, 4) = (6-3) = 9 G, 5) = (7-3) = 6 G, )) = G, 3)) = MINDIST(Q,) = = 5. Note that MINDIST(Q,) lower bouds D(Q,A) = ( 5 4) + (3 6) + (5 ) + (6 ) + (7 ) = 9.37 ad D(Q,B)= = ( 5 4) + (3 3) + (5 5) + (6 ) + (7 3) = 6.48 (formal proof below) Lemma4: MINDIST(Q,) lower bouds D(Q,C) for ay tme seres C uder U. Proof: We wll frst show MINDIST(Q,,t) lower bouds D(Q,C,t) = (q t -c t ) for ay tme seres C uder U. We kow that c t must le oe of the actve regos (Lemma 3). Wthout loss of geeralty, let us assume that c t les a actve rego G.e. G[] c t G[3]. Hece MINDIST(Q,G,t) D(Q,C,t). Also, MINDIST(Q,,t) <= MINDIST(Q,G,t) (by defto of MINDIST(Q,,t)). Hece MINDIST(Q,,t) lower bouds D(Q,C,t). Sce MINDIST(Q,) = t = MINDIST ( Q,, t) ad D(Q,C) = MINDIST( Q, C, t), t = MINDIST(Q,,t) D(Q,C,t) mples MINDIST(Q,) D(Q,C). Note that, geeral, lower the umber of actve regos at ay stat of tme, hgher the MINDIST, better the performace of the K-NN algorthm. Also, arrower the regos alog the value dmeso, hgher the MINDIST. The above two prcples justfy our choce of the dmesos of the APCA space. The odd dmesos help clusterg APCA pots wth smlar cv s, thus keepg the regos arrow alog the value dmeso. The eve dmesos help clusterg APCA pots that are approxmately alged at the segmet ed pots, thus esurg oly oe rego (mmum possble) s actve for most stats of tme. Algorthm ExactageSearch(Q, ε, T) beg. f T s a o-leaf ode. for each chld U of T 3. f MINDIST(Q,) ε ExactageSearch(Q, ε, U); // s MB of U 4. edfor 5. else // T s a leaf ode 6. for each APCA pot C T 7. f D LB (Q,C) ε 8. etreve full tme seres C from database; 9. f D(Q,C) ε Add C to result;. edf. edfor. edf ed Table 7:age search algorthm to retreve all the tme seres wth a rage of ε from query tme seres Q. The fucto s voked as ExactageSearch(Q, ε, root_ode_of_dex). Although we have focussed o K-NN search ths secto, the deftos of D LB ad MINDIST proposed ths paper are also eeded for aswerg rage queres usg a multdmesoal dex structure. The rage search algorthm s show Table 7. It s a straghtforward -tree-style recursve search algorthm combed wth the GEMINI rage query

Preprocess a planar map S. Given a query point p, report the face of S containing p. Goal: O(n)-size data structure that enables O(log n) query time.

Preprocess a planar map S. Given a query point p, report the face of S containing p. Goal: O(n)-size data structure that enables O(log n) query time. Computatoal Geometry Chapter 6 Pot Locato 1 Problem Defto Preprocess a plaar map S. Gve a query pot p, report the face of S cotag p. S Goal: O()-sze data structure that eables O(log ) query tme. C p E

More information

6.7 Network analysis. 6.7.1 Introduction. References - Network analysis. Topological analysis

6.7 Network analysis. 6.7.1 Introduction. References - Network analysis. Topological analysis 6.7 Network aalyss Le data that explctly store topologcal formato are called etwork data. Besdes spatal operatos, several methods of spatal aalyss are applcable to etwork data. Fgure: Network data Refereces

More information

APPENDIX III THE ENVELOPE PROPERTY

APPENDIX III THE ENVELOPE PROPERTY Apped III APPENDIX III THE ENVELOPE PROPERTY Optmzato mposes a very strog structure o the problem cosdered Ths s the reaso why eoclasscal ecoomcs whch assumes optmzg behavour has bee the most successful

More information

Chapter Eight. f : R R

Chapter Eight. f : R R Chapter Eght f : R R 8. Itroducto We shall ow tur our atteto to the very mportat specal case of fuctos that are real, or scalar, valued. These are sometmes called scalar felds. I the very, but mportat,

More information

Optimal multi-degree reduction of Bézier curves with constraints of endpoints continuity

Optimal multi-degree reduction of Bézier curves with constraints of endpoints continuity Computer Aded Geometrc Desg 19 (2002 365 377 wwwelsevercom/locate/comad Optmal mult-degree reducto of Bézer curves wth costrats of edpots cotuty Guo-Dog Che, Guo-J Wag State Key Laboratory of CAD&CG, Isttute

More information

1. The Time Value of Money

1. The Time Value of Money Corporate Face [00-0345]. The Tme Value of Moey. Compoudg ad Dscoutg Captalzato (compoudg, fdg future values) s a process of movg a value forward tme. It yelds the future value gve the relevat compoudg

More information

STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS. x, where. = y - ˆ " 1

STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS. x, where. = y - ˆ  1 STATISTICAL PROPERTIES OF LEAST SQUARES ESTIMATORS Recall Assumpto E(Y x) η 0 + η x (lear codtoal mea fucto) Data (x, y ), (x 2, y 2 ),, (x, y ) Least squares estmator ˆ E (Y x) ˆ " 0 + ˆ " x, where ˆ

More information

The analysis of annuities relies on the formula for geometric sums: r k = rn+1 1 r 1. (2.1) k=0

The analysis of annuities relies on the formula for geometric sums: r k = rn+1 1 r 1. (2.1) k=0 Chapter 2 Autes ad loas A auty s a sequece of paymets wth fxed frequecy. The term auty orgally referred to aual paymets (hece the ame), but t s ow also used for paymets wth ay frequecy. Autes appear may

More information

Abraham Zaks. Technion I.I.T. Haifa ISRAEL. and. University of Haifa, Haifa ISRAEL. Abstract

Abraham Zaks. Technion I.I.T. Haifa ISRAEL. and. University of Haifa, Haifa ISRAEL. Abstract Preset Value of Autes Uder Radom Rates of Iterest By Abraham Zas Techo I.I.T. Hafa ISRAEL ad Uversty of Hafa, Hafa ISRAEL Abstract Some attempts were made to evaluate the future value (FV) of the expected

More information

IDENTIFICATION OF THE DYNAMICS OF THE GOOGLE S RANKING ALGORITHM. A. Khaki Sedigh, Mehdi Roudaki

IDENTIFICATION OF THE DYNAMICS OF THE GOOGLE S RANKING ALGORITHM. A. Khaki Sedigh, Mehdi Roudaki IDENIFICAION OF HE DYNAMICS OF HE GOOGLE S RANKING ALGORIHM A. Khak Sedgh, Mehd Roudak Cotrol Dvso, Departmet of Electrcal Egeerg, K.N.oos Uversty of echology P. O. Box: 16315-1355, ehra, Ira sedgh@eetd.ktu.ac.r,

More information

ANOVA Notes Page 1. Analysis of Variance for a One-Way Classification of Data

ANOVA Notes Page 1. Analysis of Variance for a One-Way Classification of Data ANOVA Notes Page Aalss of Varace for a Oe-Wa Classfcato of Data Cosder a sgle factor or treatmet doe at levels (e, there are,, 3, dfferet varatos o the prescrbed treatmet) Wth a gve treatmet level there

More information

Fractal-Structured Karatsuba`s Algorithm for Binary Field Multiplication: FK

Fractal-Structured Karatsuba`s Algorithm for Binary Field Multiplication: FK Fractal-Structured Karatsuba`s Algorthm for Bary Feld Multplcato: FK *The authors are worg at the Isttute of Mathematcs The Academy of Sceces of DPR Korea. **Address : U Jog dstrct Kwahadog Number Pyogyag

More information

Numerical Methods with MS Excel

Numerical Methods with MS Excel TMME, vol4, o.1, p.84 Numercal Methods wth MS Excel M. El-Gebely & B. Yushau 1 Departmet of Mathematcal Sceces Kg Fahd Uversty of Petroleum & Merals. Dhahra, Saud Araba. Abstract: I ths ote we show how

More information

The Digital Signature Scheme MQQ-SIG

The Digital Signature Scheme MQQ-SIG The Dgtal Sgature Scheme MQQ-SIG Itellectual Property Statemet ad Techcal Descrpto Frst publshed: 10 October 2010, Last update: 20 December 2010 Dalo Glgorosk 1 ad Rue Stesmo Ødegård 2 ad Rue Erled Jese

More information

Average Price Ratios

Average Price Ratios Average Prce Ratos Morgstar Methodology Paper August 3, 2005 2005 Morgstar, Ic. All rghts reserved. The formato ths documet s the property of Morgstar, Ic. Reproducto or trascrpto by ay meas, whole or

More information

Applications of Support Vector Machine Based on Boolean Kernel to Spam Filtering

Applications of Support Vector Machine Based on Boolean Kernel to Spam Filtering Moder Appled Scece October, 2009 Applcatos of Support Vector Mache Based o Boolea Kerel to Spam Flterg Shugag Lu & Keb Cu School of Computer scece ad techology, North Cha Electrc Power Uversty Hebe 071003,

More information

SHAPIRO-WILK TEST FOR NORMALITY WITH KNOWN MEAN

SHAPIRO-WILK TEST FOR NORMALITY WITH KNOWN MEAN SHAPIRO-WILK TEST FOR NORMALITY WITH KNOWN MEAN Wojcech Zelńsk Departmet of Ecoometrcs ad Statstcs Warsaw Uversty of Lfe Sceces Nowoursyowska 66, -787 Warszawa e-mal: wojtekzelsk@statystykafo Zofa Hausz,

More information

Online Appendix: Measured Aggregate Gains from International Trade

Online Appendix: Measured Aggregate Gains from International Trade Ole Appedx: Measured Aggregate Gas from Iteratoal Trade Arel Burste UCLA ad NBER Javer Cravo Uversty of Mchga March 3, 2014 I ths ole appedx we derve addtoal results dscussed the paper. I the frst secto,

More information

DECISION MAKING WITH THE OWA OPERATOR IN SPORT MANAGEMENT

DECISION MAKING WITH THE OWA OPERATOR IN SPORT MANAGEMENT ESTYLF08, Cuecas Meras (Meres - Lagreo), 7-9 de Septembre de 2008 DECISION MAKING WITH THE OWA OPERATOR IN SPORT MANAGEMENT José M. Mergó Aa M. Gl-Lafuete Departmet of Busess Admstrato, Uversty of Barceloa

More information

Chapter 3. AMORTIZATION OF LOAN. SINKING FUNDS R =

Chapter 3. AMORTIZATION OF LOAN. SINKING FUNDS R = Chapter 3. AMORTIZATION OF LOAN. SINKING FUNDS Objectves of the Topc: Beg able to formalse ad solve practcal ad mathematcal problems, whch the subjects of loa amortsato ad maagemet of cumulatve fuds are

More information

On Error Detection with Block Codes

On Error Detection with Block Codes BULGARIAN ACADEMY OF SCIENCES CYBERNETICS AND INFORMATION TECHNOLOGIES Volume 9, No 3 Sofa 2009 O Error Detecto wth Block Codes Rostza Doduekova Chalmers Uversty of Techology ad the Uversty of Gotheburg,

More information

CHAPTER 2. Time Value of Money 6-1

CHAPTER 2. Time Value of Money 6-1 CHAPTER 2 Tme Value of Moey 6- Tme Value of Moey (TVM) Tme Les Future value & Preset value Rates of retur Autes & Perpetutes Ueve cash Flow Streams Amortzato 6-2 Tme les 0 2 3 % CF 0 CF CF 2 CF 3 Show

More information

Speeding up k-means Clustering by Bootstrap Averaging

Speeding up k-means Clustering by Bootstrap Averaging Speedg up -meas Clusterg by Bootstrap Averagg Ia Davdso ad Ashw Satyaarayaa Computer Scece Dept, SUNY Albay, NY, USA,. {davdso, ashw}@cs.albay.edu Abstract K-meas clusterg s oe of the most popular clusterg

More information

An Effectiveness of Integrated Portfolio in Bancassurance

An Effectiveness of Integrated Portfolio in Bancassurance A Effectveess of Itegrated Portfolo Bacassurace Taea Karya Research Ceter for Facal Egeerg Isttute of Ecoomc Research Kyoto versty Sayouu Kyoto 606-850 Japa arya@eryoto-uacp Itroducto As s well ow the

More information

The Gompertz-Makeham distribution. Fredrik Norström. Supervisor: Yuri Belyaev

The Gompertz-Makeham distribution. Fredrik Norström. Supervisor: Yuri Belyaev The Gompertz-Makeham dstrbuto by Fredrk Norström Master s thess Mathematcal Statstcs, Umeå Uversty, 997 Supervsor: Yur Belyaev Abstract Ths work s about the Gompertz-Makeham dstrbuto. The dstrbuto has

More information

We present a new approach to pricing American-style derivatives that is applicable to any Markovian setting

We present a new approach to pricing American-style derivatives that is applicable to any Markovian setting MANAGEMENT SCIENCE Vol. 52, No., Jauary 26, pp. 95 ss 25-99 ess 526-55 6 52 95 forms do.287/msc.5.447 26 INFORMS Prcg Amerca-Style Dervatves wth Europea Call Optos Scott B. Laprse BAE Systems, Advaced

More information

Dynamic Two-phase Truncated Rayleigh Model for Release Date Prediction of Software

Dynamic Two-phase Truncated Rayleigh Model for Release Date Prediction of Software J. Software Egeerg & Applcatos 3 63-69 do:.436/jsea..367 Publshed Ole Jue (http://www.scrp.org/joural/jsea) Dyamc Two-phase Trucated Raylegh Model for Release Date Predcto of Software Lafe Qa Qgchua Yao

More information

T = 1/freq, T = 2/freq, T = i/freq, T = n (number of cash flows = freq n) are :

T = 1/freq, T = 2/freq, T = i/freq, T = n (number of cash flows = freq n) are : Bullets bods Let s descrbe frst a fxed rate bod wthout amortzg a more geeral way : Let s ote : C the aual fxed rate t s a percetage N the otoal freq ( 2 4 ) the umber of coupo per year R the redempto of

More information

Group Nearest Neighbor Queries

Group Nearest Neighbor Queries Group Nearest Neghbor Queres Dmtrs Papadas Qogmao She Yufe Tao Kyrakos Mouratds Departmet of Computer Scece Hog Kog Uversty of Scece ad Techology Clear Water Bay, Hog Kog {dmtrs, qmshe, kyrakos}@cs.ust.hk

More information

A New Bayesian Network Method for Computing Bottom Event's Structural Importance Degree using Jointree

A New Bayesian Network Method for Computing Bottom Event's Structural Importance Degree using Jointree , pp.277-288 http://dx.do.org/10.14257/juesst.2015.8.1.25 A New Bayesa Network Method for Computg Bottom Evet's Structural Importace Degree usg Jotree Wag Yao ad Su Q School of Aeroautcs, Northwester Polytechcal

More information

Maintenance Scheduling of Distribution System with Optimal Economy and Reliability

Maintenance Scheduling of Distribution System with Optimal Economy and Reliability Egeerg, 203, 5, 4-8 http://dx.do.org/0.4236/eg.203.59b003 Publshed Ole September 203 (http://www.scrp.org/joural/eg) Mateace Schedulg of Dstrbuto System wth Optmal Ecoomy ad Relablty Syua Hog, Hafeg L,

More information

ADAPTATION OF SHAPIRO-WILK TEST TO THE CASE OF KNOWN MEAN

ADAPTATION OF SHAPIRO-WILK TEST TO THE CASE OF KNOWN MEAN Colloquum Bometrcum 4 ADAPTATION OF SHAPIRO-WILK TEST TO THE CASE OF KNOWN MEAN Zofa Hausz, Joaa Tarasńska Departmet of Appled Mathematcs ad Computer Scece Uversty of Lfe Sceces Lubl Akademcka 3, -95 Lubl

More information

Green Master based on MapReduce Cluster

Green Master based on MapReduce Cluster Gree Master based o MapReduce Cluster Mg-Zh Wu, Yu-Chag L, We-Tsog Lee, Yu-Su L, Fog-Hao Lu Dept of Electrcal Egeerg Tamkag Uversty, Tawa, ROC Dept of Electrcal Egeerg Tamkag Uversty, Tawa, ROC Dept of

More information

Three Dimensional Interpolation of Video Signals

Three Dimensional Interpolation of Video Signals Three Dmesoal Iterpolato of Vdeo Sgals Elham Shahfard March 0 th 006 Outle A Bref reve of prevous tals Dgtal Iterpolato Bascs Upsamplg D Flter Desg Issues Ifte Impulse Respose Fte Impulse Respose Desged

More information

ECONOMIC CHOICE OF OPTIMUM FEEDER CABLE CONSIDERING RISK ANALYSIS. University of Brasilia (UnB) and The Brazilian Regulatory Agency (ANEEL), Brazil

ECONOMIC CHOICE OF OPTIMUM FEEDER CABLE CONSIDERING RISK ANALYSIS. University of Brasilia (UnB) and The Brazilian Regulatory Agency (ANEEL), Brazil ECONOMIC CHOICE OF OPTIMUM FEEDER CABE CONSIDERING RISK ANAYSIS I Camargo, F Fgueredo, M De Olvera Uversty of Brasla (UB) ad The Brazla Regulatory Agecy (ANEE), Brazl The choce of the approprate cable

More information

Statistical Pattern Recognition (CE-725) Department of Computer Engineering Sharif University of Technology

Statistical Pattern Recognition (CE-725) Department of Computer Engineering Sharif University of Technology I The Name of God, The Compassoate, The ercful Name: Problems' eys Studet ID#:. Statstcal Patter Recogto (CE-725) Departmet of Computer Egeerg Sharf Uversty of Techology Fal Exam Soluto - Sprg 202 (50

More information

A Study of Unrelated Parallel-Machine Scheduling with Deteriorating Maintenance Activities to Minimize the Total Completion Time

A Study of Unrelated Parallel-Machine Scheduling with Deteriorating Maintenance Activities to Minimize the Total Completion Time Joural of Na Ka, Vol. 0, No., pp.5-9 (20) 5 A Study of Urelated Parallel-Mache Schedulg wth Deteroratg Mateace Actvtes to Mze the Total Copleto Te Suh-Jeq Yag, Ja-Yuar Guo, Hs-Tao Lee Departet of Idustral

More information

of the relationship between time and the value of money.

of the relationship between time and the value of money. TIME AND THE VALUE OF MONEY Most agrbusess maagers are famlar wth the terms compoudg, dscoutg, auty, ad captalzato. That s, most agrbusess maagers have a tutve uderstadg that each term mples some relatoshp

More information

Models for Selecting an ERP System with Intuitionistic Trapezoidal Fuzzy Information

Models for Selecting an ERP System with Intuitionistic Trapezoidal Fuzzy Information JOURNAL OF SOFWARE, VOL 5, NO 3, MARCH 00 75 Models for Selectg a ERP System wth Itutostc rapezodal Fuzzy Iformato Guwu We, Ru L Departmet of Ecoomcs ad Maagemet, Chogqg Uversty of Arts ad Sceces, Yogchua,

More information

Network dimensioning for elastic traffic based on flow-level QoS

Network dimensioning for elastic traffic based on flow-level QoS Network dmesog for elastc traffc based o flow-level QoS 1(10) Network dmesog for elastc traffc based o flow-level QoS Pas Lassla ad Jorma Vrtamo Networkg Laboratory Helsk Uversty of Techology Itroducto

More information

RQM: A new rate-based active queue management algorithm

RQM: A new rate-based active queue management algorithm : A ew rate-based actve queue maagemet algorthm Jeff Edmods, Suprakash Datta, Patrck Dymod, Kashf Al Computer Scece ad Egeerg Departmet, York Uversty, Toroto, Caada Abstract I ths paper, we propose a ew

More information

Load Balancing Control for Parallel Systems

Load Balancing Control for Parallel Systems Proc IEEE Med Symposum o New drectos Cotrol ad Automato, Chaa (Grèce),994, pp66-73 Load Balacg Cotrol for Parallel Systems Jea-Claude Heet LAAS-CNRS, 7 aveue du Coloel Roche, 3077 Toulouse, Frace E-mal

More information

RUSSIAN ROULETTE AND PARTICLE SPLITTING

RUSSIAN ROULETTE AND PARTICLE SPLITTING RUSSAN ROULETTE AND PARTCLE SPLTTNG M. Ragheb 3/7/203 NTRODUCTON To stuatos are ecoutered partcle trasport smulatos:. a multplyg medum, a partcle such as a eutro a cosmc ray partcle or a photo may geerate

More information

Bayesian Network Representation

Bayesian Network Representation Readgs: K&F 3., 3.2, 3.3, 3.4. Bayesa Network Represetato Lecture 2 Mar 30, 20 CSE 55, Statstcal Methods, Sprg 20 Istructor: Su-I Lee Uversty of Washgto, Seattle Last tme & today Last tme Probablty theory

More information

Integrating Production Scheduling and Maintenance: Practical Implications

Integrating Production Scheduling and Maintenance: Practical Implications Proceedgs of the 2012 Iteratoal Coferece o Idustral Egeerg ad Operatos Maagemet Istabul, Turkey, uly 3 6, 2012 Itegratg Producto Schedulg ad Mateace: Practcal Implcatos Lath A. Hadd ad Umar M. Al-Turk

More information

Fast, Secure Encryption for Indexing in a Column-Oriented DBMS

Fast, Secure Encryption for Indexing in a Column-Oriented DBMS Fast, Secure Ecrypto for Idexg a Colum-Oreted DBMS Tgja Ge, Sta Zdok Brow Uversty {tge, sbz}@cs.brow.edu Abstract Networked formato systems requre strog securty guaratees because of the ew threats that

More information

10.5 Future Value and Present Value of a General Annuity Due

10.5 Future Value and Present Value of a General Annuity Due Chapter 10 Autes 371 5. Thomas leases a car worth $4,000 at.99% compouded mothly. He agrees to make 36 lease paymets of $330 each at the begg of every moth. What s the buyout prce (resdual value of the

More information

ANALYTICAL MODEL FOR TCP FILE TRANSFERS OVER UMTS. Janne Peisa Ericsson Research 02420 Jorvas, Finland. Michael Meyer Ericsson Research, Germany

ANALYTICAL MODEL FOR TCP FILE TRANSFERS OVER UMTS. Janne Peisa Ericsson Research 02420 Jorvas, Finland. Michael Meyer Ericsson Research, Germany ANALYTICAL MODEL FOR TCP FILE TRANSFERS OVER UMTS Jae Pesa Erco Research 4 Jorvas, Flad Mchael Meyer Erco Research, Germay Abstract Ths paper proposes a farly complex model to aalyze the performace of

More information

Constrained Cubic Spline Interpolation for Chemical Engineering Applications

Constrained Cubic Spline Interpolation for Chemical Engineering Applications Costraed Cubc Sple Iterpolato or Chemcal Egeerg Applcatos b CJC Kruger Summar Cubc sple terpolato s a useul techque to terpolate betwee kow data pots due to ts stable ad smooth characterstcs. Uortuatel

More information

Efficient Traceback of DoS Attacks using Small Worlds in MANET

Efficient Traceback of DoS Attacks using Small Worlds in MANET Effcet Traceback of DoS Attacks usg Small Worlds MANET Yog Km, Vshal Sakhla, Ahmed Helmy Departmet. of Electrcal Egeerg, Uversty of Souther Calfora, U.S.A {yogkm, sakhla, helmy}@ceg.usc.edu Abstract Moble

More information

n. We know that the sum of squares of p independent standard normal variables has a chi square distribution with p degrees of freedom.

n. We know that the sum of squares of p independent standard normal variables has a chi square distribution with p degrees of freedom. UMEÅ UNIVERSITET Matematsk-statstska sttutoe Multvarat dataaalys för tekologer MSTB0 PA TENTAMEN 004-0-9 LÖSNINGSFÖRSLAG TILL TENTAMEN I MATEMATISK STATISTIK Multvarat dataaalys för tekologer B, 5 poäg.

More information

Optimal replacement and overhaul decisions with imperfect maintenance and warranty contracts

Optimal replacement and overhaul decisions with imperfect maintenance and warranty contracts Optmal replacemet ad overhaul decsos wth mperfect mateace ad warraty cotracts R. Pascual Departmet of Mechacal Egeerg, Uversdad de Chle, Caslla 2777, Satago, Chle Phoe: +56-2-6784591 Fax:+56-2-689657 rpascual@g.uchle.cl

More information

Security Analysis of RAPP: An RFID Authentication Protocol based on Permutation

Security Analysis of RAPP: An RFID Authentication Protocol based on Permutation Securty Aalyss of RAPP: A RFID Authetcato Protocol based o Permutato Wag Shao-hu,,, Ha Zhje,, Lu Sujua,, Che Da-we, {College of Computer, Najg Uversty of Posts ad Telecommucatos, Najg 004, Cha Jagsu Hgh

More information

Robust Realtime Face Recognition And Tracking System

Robust Realtime Face Recognition And Tracking System JCS& Vol. 9 No. October 9 Robust Realtme Face Recogto Ad rackg System Ka Che,Le Ju Zhao East Cha Uversty of Scece ad echology Emal:asa85@hotmal.com Abstract here s some very mportat meag the study of realtme

More information

Statistical Intrusion Detector with Instance-Based Learning

Statistical Intrusion Detector with Instance-Based Learning Iformatca 5 (00) xxx yyy Statstcal Itruso Detector wth Istace-Based Learg Iva Verdo, Boja Nova Faulteta za eletroteho raualštvo Uverza v Marboru Smetaova 7, 000 Marbor, Sloveja va.verdo@sol.et eywords:

More information

Common p-belief: The General Case

Common p-belief: The General Case GAMES AND ECONOMIC BEHAVIOR 8, 738 997 ARTICLE NO. GA97053 Commo p-belef: The Geeral Case Atsush Kaj* ad Stephe Morrs Departmet of Ecoomcs, Uersty of Pesylaa Receved February, 995 We develop belef operators

More information

Compressive Sensing over Strongly Connected Digraph and Its Application in Traffic Monitoring

Compressive Sensing over Strongly Connected Digraph and Its Application in Traffic Monitoring Compressve Sesg over Strogly Coected Dgraph ad Its Applcato Traffc Motorg Xao Q, Yogca Wag, Yuexua Wag, Lwe Xu Isttute for Iterdscplary Iformato Sceces, Tsghua Uversty, Bejg, Cha {qxao3, kyo.c}@gmal.com,

More information

Capacitated Production Planning and Inventory Control when Demand is Unpredictable for Most Items: The No B/C Strategy

Capacitated Production Planning and Inventory Control when Demand is Unpredictable for Most Items: The No B/C Strategy SCHOOL OF OPERATIONS RESEARCH AND INDUSTRIAL ENGINEERING COLLEGE OF ENGINEERING CORNELL UNIVERSITY ITHACA, NY 4853-380 TECHNICAL REPORT Jue 200 Capactated Producto Plag ad Ivetory Cotrol whe Demad s Upredctable

More information

How To Value An Annuity

How To Value An Annuity Future Value of a Auty After payg all your blls, you have $200 left each payday (at the ed of each moth) that you wll put to savgs order to save up a dow paymet for a house. If you vest ths moey at 5%

More information

Finito: A Faster, Permutable Incremental Gradient Method for Big Data Problems

Finito: A Faster, Permutable Incremental Gradient Method for Big Data Problems Fto: A Faster, Permutable Icremetal Gradet Method for Bg Data Problems Aaro J Defazo Tbéro S Caetao Just Domke NICTA ad Australa Natoal Uversty AARONDEFAZIO@ANUEDUAU TIBERIOCAETANO@NICTACOMAU JUSTINDOMKE@NICTACOMAU

More information

An IG-RS-SVM classifier for analyzing reviews of E-commerce product

An IG-RS-SVM classifier for analyzing reviews of E-commerce product Iteratoal Coferece o Iformato Techology ad Maagemet Iovato (ICITMI 205) A IG-RS-SVM classfer for aalyzg revews of E-commerce product Jaju Ye a, Hua Re b ad Hagxa Zhou c * College of Iformato Egeerg, Cha

More information

Plastic Number: Construction and Applications

Plastic Number: Construction and Applications Scet f c 0 Advaced Advaced Scetfc 0 December,.. 0 Plastc Number: Costructo ad Applcatos Lua Marohć Polytechc of Zagreb, 0000 Zagreb, Croata lua.marohc@tvz.hr Thaa Strmeč Polytechc of Zagreb, 0000 Zagreb,

More information

The Analysis of Development of Insurance Contract Premiums of General Liability Insurance in the Business Insurance Risk

The Analysis of Development of Insurance Contract Premiums of General Liability Insurance in the Business Insurance Risk The Aalyss of Developmet of Isurace Cotract Premums of Geeral Lablty Isurace the Busess Isurace Rsk the Frame of the Czech Isurace Market 1998 011 Scetfc Coferece Jue, 10. - 14. 013 Pavla Kubová Departmet

More information

Chapter 3 0.06 = 3000 ( 1.015 ( 1 ) Present Value of an Annuity. Section 4 Present Value of an Annuity; Amortization

Chapter 3 0.06 = 3000 ( 1.015 ( 1 ) Present Value of an Annuity. Section 4 Present Value of an Annuity; Amortization Chapter 3 Mathematcs of Face Secto 4 Preset Value of a Auty; Amortzato Preset Value of a Auty I ths secto, we wll address the problem of determg the amout that should be deposted to a accout ow at a gve

More information

The impact of service-oriented architecture on the scheduling algorithm in cloud computing

The impact of service-oriented architecture on the scheduling algorithm in cloud computing Iteratoal Research Joural of Appled ad Basc Sceces 2015 Avalable ole at www.rjabs.com ISSN 2251-838X / Vol, 9 (3): 387-392 Scece Explorer Publcatos The mpact of servce-oreted archtecture o the schedulg

More information

where p is the centroid of the neighbors of p. Consider the eigenvector problem

where p is the centroid of the neighbors of p. Consider the eigenvector problem Vrtual avgato of teror structures by ldar Yogja X a, Xaolg L a, Ye Dua a, Norbert Maerz b a Uversty of Mssour at Columba b Mssour Uversty of Scece ad Techology ABSTRACT I ths project, we propose to develop

More information

A Smart Machine Vision System for PCB Inspection

A Smart Machine Vision System for PCB Inspection A Smart Mache Vso System for PCB Ispecto Te Q Che, JaX Zhag, YouNg Zhou ad Y Lu Murphey Please address all correspodece to Departmet of Electrcal ad Computer Egeerg Uversty of Mchga - Dearbor, Dearbor,

More information

IP Network Topology Link Prediction Based on Improved Local Information Similarity Algorithm

IP Network Topology Link Prediction Based on Improved Local Information Similarity Algorithm Iteratoal Joural of Grd Dstrbuto Computg, pp.141-150 http://dx.do.org/10.14257/jgdc.2015.8.6.14 IP Network Topology Lk Predcto Based o Improved Local Iformato mlarty Algorthm Che Yu* 1, 2 ad Dua Zhem 1

More information

Impact of Interference on the GPRS Multislot Link Level Performance

Impact of Interference on the GPRS Multislot Link Level Performance Impact of Iterferece o the GPRS Multslot Lk Level Performace Javer Gozalvez ad Joh Dulop Uversty of Strathclyde - Departmet of Electroc ad Electrcal Egeerg - George St - Glasgow G-XW- Scotlad Ph.: + 8

More information

Relaxation Methods for Iterative Solution to Linear Systems of Equations

Relaxation Methods for Iterative Solution to Linear Systems of Equations Relaxato Methods for Iteratve Soluto to Lear Systems of Equatos Gerald Recktewald Portlad State Uversty Mechacal Egeerg Departmet gerry@me.pdx.edu Prmary Topcs Basc Cocepts Statoary Methods a.k.a. Relaxato

More information

Credibility Premium Calculation in Motor Third-Party Liability Insurance

Credibility Premium Calculation in Motor Third-Party Liability Insurance Advaces Mathematcal ad Computatoal Methods Credblty remum Calculato Motor Thrd-arty Lablty Isurace BOHA LIA, JAA KUBAOVÁ epartmet of Mathematcs ad Quattatve Methods Uversty of ardubce Studetská 95, 53

More information

Performance Attribution. Methodology Overview

Performance Attribution. Methodology Overview erformace Attrbuto Methodology Overvew Faba SUAREZ March 2004 erformace Attrbuto Methodology 1.1 Itroducto erformace Attrbuto s a set of techques that performace aalysts use to expla why a portfolo's performace

More information

An Approach to Evaluating the Computer Network Security with Hesitant Fuzzy Information

An Approach to Evaluating the Computer Network Security with Hesitant Fuzzy Information A Approach to Evaluatg the Computer Network Securty wth Hestat Fuzzy Iformato Jafeg Dog A Approach to Evaluatg the Computer Network Securty wth Hestat Fuzzy Iformato Jafeg Dog, Frst ad Correspodg Author

More information

Reinsurance and the distribution of term insurance claims

Reinsurance and the distribution of term insurance claims Resurace ad the dstrbuto of term surace clams By Rchard Bruyel FIAA, FNZSA Preseted to the NZ Socety of Actuares Coferece Queestow - November 006 1 1 Itroducto Ths paper vestgates the effect of resurace

More information

A particle swarm optimization to vehicle routing problem with fuzzy demands

A particle swarm optimization to vehicle routing problem with fuzzy demands A partcle swarm optmzato to vehcle routg problem wth fuzzy demads Yag Peg, Ye-me Qa A partcle swarm optmzato to vehcle routg problem wth fuzzy demads Yag Peg 1,Ye-me Qa 1 School of computer ad formato

More information

The Time Value of Money

The Time Value of Money The Tme Value of Moey 1 Iversemet Optos Year: 1624 Property Traded: Mahatta Islad Prce : $24.00, FV of $24 @ 6%: FV = $24 (1+0.06) 388 = $158.08 bllo Opto 1 0 1 2 3 4 5 t ($519.37) 0 0 0 0 $1,000 Opto

More information

Classic Problems at a Glance using the TVM Solver

Classic Problems at a Glance using the TVM Solver C H A P T E R 2 Classc Problems at a Glace usg the TVM Solver The table below llustrates the most commo types of classc face problems. The formulas are gve for each calculato. A bref troducto to usg the

More information

Lecture 7. Norms and Condition Numbers

Lecture 7. Norms and Condition Numbers Lecture 7 Norms ad Codto Numbers To dscuss the errors umerca probems vovg vectors, t s usefu to empo orms. Vector Norm O a vector space V, a orm s a fucto from V to the set of o-egatve reas that obes three

More information

Impact of Mobility Prediction on the Temporal Stability of MANET Clustering Algorithms *

Impact of Mobility Prediction on the Temporal Stability of MANET Clustering Algorithms * Impact of Moblty Predcto o the Temporal Stablty of MANET Clusterg Algorthms * Aravdha Vekateswara, Vekatesh Saraga, Nataraa Gautam 1, Ra Acharya Departmet of Comp. Sc. & Egr. Pesylvaa State Uversty Uversty

More information

Banking (Early Repayment of Housing Loans) Order, 5762 2002 1

Banking (Early Repayment of Housing Loans) Order, 5762 2002 1 akg (Early Repaymet of Housg Loas) Order, 5762 2002 y vrtue of the power vested me uder Secto 3 of the akg Ordace 94 (hereafter, the Ordace ), followg cosultato wth the Commttee, ad wth the approval of

More information

Settlement Prediction by Spatial-temporal Random Process

Settlement Prediction by Spatial-temporal Random Process Safety, Relablty ad Rs of Structures, Ifrastructures ad Egeerg Systems Furuta, Fragopol & Shozua (eds Taylor & Fracs Group, Lodo, ISBN 978---77- Settlemet Predcto by Spatal-temporal Radom Process P. Rugbaapha

More information

Near Neighbor Distribution in Sets of Fractal Nature

Near Neighbor Distribution in Sets of Fractal Nature Iteratoal Joural of Computer Iformato Systems ad Idustral Maagemet Applcatos. ISS 250-7988 Volume 5 (202) 3 pp. 59-66 MIR Labs, www.mrlabs.et/jcsm/dex.html ear eghbor Dstrbuto Sets of Fractal ature Marcel

More information

On Savings Accounts in Semimartingale Term Structure Models

On Savings Accounts in Semimartingale Term Structure Models O Savgs Accouts Semmartgale Term Structure Models Frak Döberle Mart Schwezer moeyshelf.com Techsche Uverstät Berl Bockehemer Ladstraße 55 Fachberech Mathematk, MA 7 4 D 6325 Frakfurt am Ma Straße des 17.

More information

Optimal Packetization Interval for VoIP Applications Over IEEE 802.16 Networks

Optimal Packetization Interval for VoIP Applications Over IEEE 802.16 Networks Optmal Packetzato Iterval for VoIP Applcatos Over IEEE 802.16 Networks Sheha Perera Harsha Srsea Krzysztof Pawlkowsk Departmet of Electrcal & Computer Egeerg Uversty of Caterbury New Zealad sheha@elec.caterbury.ac.z

More information

USEFULNESS OF BOOTSTRAPPING IN PORTFOLIO MANAGEMENT

USEFULNESS OF BOOTSTRAPPING IN PORTFOLIO MANAGEMENT USEFULNESS OF BOOTSTRAPPING IN PORTFOLIO MANAGEMENT Radovaov Bors Faculty of Ecoomcs Subotca Segedsk put 9-11 Subotca 24000 E-mal: radovaovb@ef.us.ac.rs Marckć Aleksadra Faculty of Ecoomcs Subotca Segedsk

More information

A Single Machine Scheduling with Periodic Maintenance

A Single Machine Scheduling with Periodic Maintenance A Sgle Mache Schedulg wth Perodc Mateace Fracsco Ágel-Bello Ada Álvarez 2 Joaquí Pacheco 3 Irs Martíez Ceter for Qualty ad Maufacturg, Tecológco de Moterrey, Eugeo Garza Sada 250, 64849 Moterrey, NL, Meco

More information

RESEARCH ON PERFORMANCE MODELING OF TRANSACTIONAL CLOUD APPLICATIONS

RESEARCH ON PERFORMANCE MODELING OF TRANSACTIONAL CLOUD APPLICATIONS Joural of Theoretcal ad Appled Iformato Techology 3 st October 22. Vol. 44 No.2 25-22 JATIT & LLS. All rghts reserved. ISSN: 992-8645 www.jatt.org E-ISSN: 87-395 RESEARCH ON PERFORMANCE MODELING OF TRANSACTIONAL

More information

Projection model for Computer Network Security Evaluation with interval-valued intuitionistic fuzzy information. Qingxiang Li

Projection model for Computer Network Security Evaluation with interval-valued intuitionistic fuzzy information. Qingxiang Li Iteratoal Joural of Scece Vol No7 05 ISSN: 83-4890 Proecto model for Computer Network Securty Evaluato wth terval-valued tutostc fuzzy formato Qgxag L School of Software Egeerg Chogqg Uversty of rts ad

More information

CH. V ME256 STATICS Center of Gravity, Centroid, and Moment of Inertia CENTER OF GRAVITY AND CENTROID

CH. V ME256 STATICS Center of Gravity, Centroid, and Moment of Inertia CENTER OF GRAVITY AND CENTROID CH. ME56 STTICS Ceter of Gravt, Cetrod, ad Momet of Ierta CENTE OF GITY ND CENTOID 5. CENTE OF GITY ND CENTE OF MSS FO SYSTEM OF PTICES Ceter of Gravt. The ceter of gravt G s a pot whch locates the resultat

More information

An Evaluation of Naïve Bayesian Anti-Spam Filtering Techniques

An Evaluation of Naïve Bayesian Anti-Spam Filtering Techniques Proceedgs of the 2007 IEEE Workshop o Iformato Assurace Uted tates Mltary Academy, West Pot, Y 20-22 Jue 2007 A Evaluato of aïve Bayesa At-pam Flterg Techques Vkas P. Deshpade, Robert F. Erbacher, ad Chrs

More information

Simple Linear Regression

Simple Linear Regression Smple Lear Regresso Regresso equato a equato that descrbes the average relatoshp betwee a respose (depedet) ad a eplaator (depedet) varable. 6 8 Slope-tercept equato for a le m b (,6) slope. (,) 6 6 8

More information

Automated Event Registration System in Corporation

Automated Event Registration System in Corporation teratoal Joural of Advaces Computer Scece ad Techology JACST), Vol., No., Pages : 0-0 0) Specal ssue of CACST 0 - Held durg 09-0 May, 0 Malaysa Automated Evet Regstrato System Corporato Zafer Al-Makhadmee

More information

Polyphase Filters. Section 12.4 Porat 1/39

Polyphase Filters. Section 12.4 Porat 1/39 Polyphase Flters Secto.4 Porat /39 .4 Polyphase Flters Polyphase s a way of dog saplg-rate coverso that leads to very effcet pleetatos. But ore tha that, t leads to very geeral vewpots that are useful

More information

STOCHASTIC approximation algorithms have several

STOCHASTIC approximation algorithms have several IEEE TRANSACTIONS ON INFORMATION THEORY, VOL 60, NO 10, OCTOBER 2014 6609 Trackg a Markov-Modulated Statoary Degree Dstrbuto of a Dyamc Radom Graph Mazyar Hamd, Vkram Krshamurthy, Fellow, IEEE, ad George

More information

MDM 4U PRACTICE EXAMINATION

MDM 4U PRACTICE EXAMINATION MDM 4U RCTICE EXMINTION Ths s a ractce eam. It does ot cover all the materal ths course ad should ot be the oly revew that you do rearato for your fal eam. Your eam may cota questos that do ot aear o ths

More information

Using Phase Swapping to Solve Load Phase Balancing by ADSCHNN in LV Distribution Network

Using Phase Swapping to Solve Load Phase Balancing by ADSCHNN in LV Distribution Network Iteratoal Joural of Cotrol ad Automato Vol.7, No.7 (204), pp.-4 http://dx.do.org/0.4257/jca.204.7.7.0 Usg Phase Swappg to Solve Load Phase Balacg by ADSCHNN LV Dstrbuto Network Chu-guo Fe ad Ru Wag College

More information

Fault Tree Analysis of Software Reliability Allocation

Fault Tree Analysis of Software Reliability Allocation Fault Tree Aalyss of Software Relablty Allocato Jawe XIANG, Kokch FUTATSUGI School of Iformato Scece, Japa Advaced Isttute of Scece ad Techology - Asahda, Tatsuokuch, Ishkawa, 92-292 Japa ad Yaxag HE Computer

More information

The Application of Intuitionistic Fuzzy Set TOPSIS Method in Employee Performance Appraisal

The Application of Intuitionistic Fuzzy Set TOPSIS Method in Employee Performance Appraisal Vol.8, No.3 (05), pp.39-344 http://dx.do.org/0.457/uesst.05.8.3.3 The pplcato of Itutostc Fuzzy Set TOPSIS Method Employee Performace pprasal Wag Yghu ad L Welu * School of Ecoomcs ad Maagemet, Shazhuag

More information

M. Salahi, F. Mehrdoust, F. Piri. CVaR Robust Mean-CVaR Portfolio Optimization

M. Salahi, F. Mehrdoust, F. Piri. CVaR Robust Mean-CVaR Portfolio Optimization M. Salah, F. Mehrdoust, F. Pr Uversty of Gula, Rasht, Ira CVaR Robust Mea-CVaR Portfolo Optmzato Abstract: Oe of the most mportat problems faced by every vestor s asset allocato. A vestor durg makg vestmet

More information

Analysis of Multi-product Break-even with Uncertain Information*

Analysis of Multi-product Break-even with Uncertain Information* Aalyss o Mult-product Break-eve wth Ucerta Iormato* Lazzar Lusa L. - Morñgo María Slva Facultad de Cecas Ecoómcas Uversdad de Bueos Ares 222 Córdoba Ave. 2 d loor C20AAQ Bueos Ares - Argeta lazzar@eco.uba.ar

More information