KronoMiner: Using Multi-Foci Navigation for the Visual Exploration of Time-Series Data

Transcription

1 CHI 2011 Session: Visul Anlytis KronoMiner: Using Multi-Foi Nvigtion for the Visul Explortion of Time-Series Dt Jin Zho1 Fnny Chevlier2 1 Deprtment of Computer Siene University of Toronto {jin rvin}@gp.toronto.eu Rvin Blkrishnn1 2 OCAD University Toronto fhevlier@o. g e f Figure 1. Explortion of four stok mrket tsets using KronoMiner. Left: the visuliztion winow of KronoMiner inluing () sttus r, () min irulr tree winow () timeline overview, () four plots omprison re, isplying (e) MgiAnlytis Lens etile view. A tooltip (f) provies itionl informtion on emn for eh segment. Both generl n ontext-se opertions re ville y invoking (g) ontext menu. ABSTRACT INTRODUCTION The nee for pttern isovery in long time-series t le reserhers to evelop intertive visuliztion tools n nlytil lgorithms for gining insight into the t. Most of the literture on time-series t visuliztion either fous on smll numer of tsks or speifi omin. We propose KronoMiner, tool tht emes new intertion n visuliztion tehniques s well s nlytil pilities for the visul explortion of time-series t. The interfe s esign hs een itertively refine se on feek from expert users. Qulittive evlution with n expert user not involve in the esign proess inites tht our prototype is promising for further reserh. The nee to nlyze time-series t is prevlent in vrious pplition omins rnging from usiness plnning to intensive re meiine. Intertive visuliztion hs proven to e n effetive pproh to nlyze time-series, ssisting otors in ignosis [25], helping omputer network engineers optimize nwith usge [22], iing finne experts mrket investments [5] or helping homeln seurity investigtions [6]. ACM Clssifition Keywors H.5.2 User Interfes: [Intertion Styles] Generl Terms Design, Humn Ftors Permission to mke igitl or hr opies of ll or prt of this work for personl or lssroom use is grnte without fee provie tht opies re not me or istriute for profit or ommeril vntge n tht opies er this notie n the full ittion on the first pge. To opy otherwise, or repulish, to post on servers or to reistriute to lists, requires prior speifi permission n/or fee. CHI 2011, My 7 12, 2011, Vnouver, BC, Cn. Copyright 2011 ACM /11/05...$ Existing intertive time-series visuliztion tools re either omin-speifi or multi-purpose. Domin-speifi tools re typilly esigne with omin experts n re rih in speilize funtionlity. They re powerful for nswering preise questions (e.g. wht is the impt of meil tretment? [10]) ut they nnot e esily pplie to other omins. They lso onfine users into preefine tsks n proeures, thus offering little flexiility in the explortion proess n leving little room for opportunisti isoveries. In ontrst, multipurpose tools re not restrite to speifi omin n thus trget wier uiene. However, they usully only support smll numer of tsks (e.g. ientifying perioiity [33]) n thus require other tools to e use in onjuntion for omplementry ut neessry explortion pilities.

2 There is therefore nee for intertive time-series t visuliztion tools tht re funtionl enough to support omplex tsks while eing flexile enough for promoting opportunisti explortion in omin-inepenent mnner. Not only oes omin-inepenene mke the tool essile to wie uiene, ut it lso llows ross-exmintion of t from multiple omins n opens up new possiilities for ollortion etween nlysts with omplementry expertise. To ress this nee we propose KronoMiner, multipurpose time-series explortion tool proviing rih nvigtion pilities n nlytil support. KronoMiner is se on ril isply tht n e rille into etils y fousing on ifferent piees of the t n rerrnging them in file mnner. The esign of KronoMiner hs een refine se on expert users feek t severl stges of the evelopment. Our ontriutions inlue (i) txonomy of esign requirements gthere from experts nlyzing time-series from vrious pplition omins, (ii) review of the esign spe of lyouts for visulizing multiple multivrite time-series, (iii) the itertive esign of KronoMiner s emonstrtion of the new visuliztion n intertion tehniques in semless tool, n (iv) expert user feek on our prototype. RELATED WORK Visul Explortion is the proess of intertively rowsing through ifferent portions of tset to gin etter unerstning of it. Not only oes Visul Explortion help for hypothesis onfirmtion, it is lso ruil for fining the unexpete n for rising new questions, espeilly when the users hve no or only vgue hypotheses out the t [31]. Although time-series t is n re of extensive stuy in Informtion Visuliztion reserh, existing intertive visuliztion tools for time-series t o not extensively support multipurpose Visul Explortion: they re either too speilize omin-speifi tools, nswering preise questions, or omin-inepenent ut limite to few pilities. Relevnt surveys inlue [24, 29]. The prevlent pproh for representing temporl t is rguly timeline visuliztion (e.g. LifeLines [25]) in whih t points re usully plotte long uniform time xis. Other pprohes use visul metphors to represent perioi temporl t. For instne, Clenr View [32] uses lenr isply to visulize univrite time-series ggregte on ily, weekly or monthly sis to fin interesting groupings of the t. SpirClok [7] provies ontinuous feek out upoming events y isplying them on spirl nlog lok. A similr spirl shpe time xis is use in Spirl Grph [33] to revel yles of time-series. All these tehniques require the t to exhiit regulrity, n therefore re of little utility for t without perioiity. For t without pprent yles, multi-foi intertion is require for the etile omprison of ifferent piees of the t while preserving ontext. For exmple, LiveRAC [22] n Mélnge [9] rely on streth n squish nvigtion. They result in single integrte view with visul istortion, even though Rihter et l. [26] suggest tht non-liner representtion of time hs negtive effets on the interprettion of the t. In ontrst, StkZoom [17] preserves time xis uniformity for eh single sequene of multi-resolution hierrhy ynmilly uilt y the user. However, the multiple foi views nnot e rerrnge, therey mking it iffiult to ompre regions of interest tht re spe out. Time-series visuliztion tools lso frequently integrte nlytil methos to support pttern mining. For instne PtternFiner [10] ssists in eteting user-efine event ptterns in meil reors y letting the user speify the esire sequene of events, their ttriutes n inter-event time spns. Although pttern mining methos help to gin insight into t, most of them require prior knowlege of the omin n of wht hrterizes n interesting pttern. In the ontext of omin-inepenent explortion, no speifi pttern mining methos n e esily pplie. Nevertheless, omining visuliztion with generl nlytil methos suh s the ifferene plot of two vriles, remins useful s n inition for tenttively ientifying regions of interest. In ontrst to PtternFiner, TimeSerher2 [5] relies on suh Visul Explortion proess y llowing the user to ynmilly selet timeox iretly on the visuliztion s query y exmple. When they support pttern mining, time-series intertive visuliztion tools usully present relevnt ptterns s list sorte y mthing sores to query [10, 5], or utomtilly lign them [25]. Sine they often fous on one pttern t time, these tools fil in proviing n overview of ll the reltionships within n ross time-series. An effetive wy for showing the reltionships is to expliitly link the regions tht mth to eh others using stright n urve lines. However, this n quikly le to visul lutter euse of too mny link rossings. A ommon pproh to ress this prolem is to unle the links [16] n represent the t on irulr lyout, s it is one in MizBee [23]. A survey of irulr visuliztions n e foun in [8]. To summrize, espite the unnt reserh for visulizing time-series t, s fr s we know, there is no system tht extensively supports omin-inepenent Visul Explortion of multiple multivrite tsets. Aville intertive visuliztion tools re either highly omin-epenent n limite to preefine tsks n proeures, or generl multipurpose tools tht support very few numer of tsks. With the knowlege we gine through our literture survey, we erive the esign guielines liste in the next setion we use s sis for the esign of KronoMiner. DESIGN GUIDELINES Before enumerting the esign requirements KronoMiner hs een uilt upon, we give kgroun on time-series t n isuss the esign spe of lyouts for visulizing them. Time-Series Dt Time-series re sequenes of t points mesure t suessive times, typilly t equl intervls. They onsist of rel numers (e.g. temperture logs), or events (e.g. meil histories me of reors of otor visits, loo tests, et). Time-series hve the ommon property of eing lrge, not only in the pture time perio n numer of smples, ut lso in the numer of oserve ttriutes [2]. 1738

3 The hrteristis of time-series vry wiely n re very epenent on the pplition omin. The temporl imension itself is speil imension tht hs multiple fets. Aigner et l. [2] hve liste the following riteri efining the ifferent types of time: (1) struture of time (liner vs. yli), (2) temporl primitive (time points vs. time intervls) n (3) time perspetive (orere time vs. rnhing time). In this work, we fous on non-intrinsilly-perioi liner sequenes of orere time points. Nevertheless, it is possile to generlize our frmework to intervls. One of the inherent limittions of tritionl time-series intertive visuliztion tools lies in regring time s eing n immutle n uniform imension. Depening on the tool, Visul Explortion is usully onstrine to t lest one of the following: (1) synhronism: ifferent ttriutes re usully time-ligne, mking it iffiult to ompre elye sequenes; (2) hronology: time-series t is often onsiere s orere t, for whih suessive piees nnot e rerrnge; (3) time sle onsisteny: ompring time-series simultneously t ifferent time sles is rrely supporte. For exmple, squishing or strething the time sle of server logs inepenently mkes it possile to just the visuliztion oring to response-time pity. Some users might lso look for su-ptterns or frtl ehviour [12]. Design requirements Before esigning n intertive visuliztion system for Visul Explortion of time-series t, we onute n informl stuy with experts from ifferent pplition omins to unerstn wht fetures users nee n how they generlly rry out their explortion of time-series t. Bse on their omments n the knowlege we gine through our literture survey, we erive the following esign requirements. R1: Multivrite, multiple n heterogeneous tsets. Deling with s vrie input t s possile is importnt for supporting multipurpose use. () multiple heterogeneous tsets: supporting timeseries from vrious heterogeneous soures to filitte ross-onept omprison n refrin from eing omin-epenent, () multivrite t: visuliztion n mnipultion of oth ouple n issoite ttriutes, () vriety of hrt representtions: proviing ifferent hrt representtions hoies. R2: Multiple oorinte views [2]. When user tions reeive immeite visul feek, multiple views i Visul Explortion y proviing iverse perspetives of the t. () lrge-sle overview: quiring generl ontext t ny explortion stte through n overview of the entire tsets help mintin mentl mp of the t [28], () multi-foi, temporl hierrhies: showing severl levels of etil simultneously in oorinte views n i in effetive rowsing of the t [17], () sie-y-sie omprison: sie-y-sie omprison is esier thn ompring visully spe-out hrts or rememering previously seen views [22]. R3: Intertion. Giving users s muh freeom s possile to mnipulte n query the t is ruil for Visul Explortion [2]. () iret mnipultion: physilly mnipulting grphil representtions of ojets hs the enefit of eing esy to use n lern n llows for greter system omprehension [27], () overview first, zoom & filter, etils on emn: the wiely followe Shneiermn s priniple [28] is reognize s effetive for eling with lrge tsets, () ynmi multi-foi, temporl hierrhies: rilling own into multiple regions of interest y time rushing (i.e. mnully selet t on n intertive t isply) eses omprison [17], () grouping, ggregting for t strtion: grouping severl regions of interest in single entity llows for user-efine lustering n t strtion, (f) history: mintining the history of the explortion proess is importnt to help users revisit importnt views n keep trk of their explortion [14], (e) nnotting: nnottions i for orgnizing n lower memory lo s they t s lels n reminers. R4: Anlytil methos. All our expert users ehoe tht semi-guie explortion se on sttistil nlysis shoul e onsiere [2]. () ross-onept reltionships: isplying the reltionships (links, sores) resulting from the omputtion of lssil omin-inepenent nlytil methos provies importnt lues for Visul Explortion, () est mth: fining the est mth etween series n their mthing sore llows for pttern mining [5], () ynmi omprison: isplying on-the-fly ompute informtion (e.g. ifferene plot) filittes ientifying regions of interest if immeite visul upte is supporte while mnipulting the t, () ustomiztion: llowing for user-efine opertors vois multipurpose tool to e too limite for speilize omin-epenent nlysis. Design Spe of Lyouts Here we review ifferent lyouts tht n e use to represent time-series t s omplement to Meyer n Munzner s txonomy [23]. Figure 2 epits the possile lyouts for multivrite (,) n multiple (-f) time-series tsets. Note tht n exhustive txonomy woul inlue spirl lyouts previously esrie, ut these ones re suite for t tht exhiits perioiity n re out of the sope of this pper. We istinguish etween multivrite t, whih implies the plots to e ouple, n multiple inepenent tsets (tht oul e multivrite or not). Multivrite t n e overli in single plot (Fig. 2), ut visul lutter might e overwhelming; n lterntive is to stk multiple hrts (Fig. 2). Different hrt types n e set for eh iniviul plot inepenently without overlp, ut this requires more sreen rel estte n might split the user s visul ttention. 1739

4 Overplotte Stke Cirulr Liner Stke e Sequentil Figure 2. Design spe of lyouts for multiple multivrite time-series. - multivrite lyouts, -f multiple tsets lyouts For plotting more thn one time-series tset t time, one n hoose etween liner n irulr lyout, n use stke or sequentil plots (Fig. 2). Liner lyouts re more fmilir to users n o not istort the t. Cirulr lyouts re more pproprite for visulizing hierrhil strutures n reltionships [8]. They lso suesfully pply to perioi t when spirlly-shpe, ut this rises issues when onsiering multiple series with ifferent perioiity n time intervl s the series hve to e pile up in the sme spirl. The sequentil liner view (Fig. 2f) mkes it iffiult to ompre time-ligne t, espeilly when the plots re spe out n nnot e rerrnge. The irulr stke lyout (Fig. 2) preserves time lignment, ut the sme t plotte on ifferent rings is pereive ifferently, n my use misinterprettion. As esrie next, the KronoMiner interfe omines oth irulr (Fig. 2e) n liner (Fig. 2) lyouts. THE KRONOMINER INTERFACE KronoMiner is n intertive visuliztion tool for Visul Explortion of multiple multivrite time-series tsets. Its esign relies on the forementione guielines n hs een itertively refine se on experts feek gthere t ifferent evelopment stges (see User Feek setion). KronoMiner ims to ese the ientifition n mnipultion of multiple su-piees of interest for finer nlysis, rther thn proviing single glol view of numerous long time-series to e nlyze s whole. It is se on ynmi hierrhil inspetion of the t t ifferent sles n lotions. The interfe (Fig. 1) omines severl oorinte omponents, tking full vntge of the iniviul strengths of eh lyout to effetively support ifferent tsks. Multi-Foi Hierrhy Tree Hierrhy Tree Visuliztion The min intertive explortion winow (Fig. 1) isplys irulr tree struture for presenting the hierrhy of multiple foi (R2). We hve hosen irulr sequentil lyout for the isply of the multi-foi hierrhy minly for its ompt representtion of tree strutures [8], ut lso to put seletle t within esy reh of the user [11]. Eh of the tree noes orrespons to t points within speifi rnge of prent series n is isplye s onentri r segment on whih t is plotte. The entrl ring ontins ll the ville entire tsets (R1,). Segments elonging to ifferent tsets n e esily istinguishe y the olor of their outlines. Detile informtion suh s timestmp ouns n time sle is isplye for eh segment. f A prent-hil reltionship of two segments is expliitly visulize s pir of she urves tht link one segment to the orresponing region of interest (ROI) on its prent. Dt points elonging to the ROI re uplite in the ssoite segment, s new n more etile view of the sme initil tset. To visully reinfore the hierrhil struture, ROIs re fille in with light gry, n rs on oth of the hil n prent segments re olor-oe to give wreness of whih portion of this t piee resies in the series. A istinguishle hue is mppe to eh tset. The she links s well s the lower ounry of eh hil segment re olore with the sme hue s its prent. The luminne onveys the position in the prent series (the righter the sooner in the prent). Moreover, when hovering over segment, its nestors n esennts in the hierrhy n the ssoite ROI in the prent re emphsize y glow effets (Fig. 1). We lso inite its lotion in the timeline overview (R2, Fig. 1), n tooltip showing etile informtion is poppe-up (R3, Fig. 1f). Duplite t is therey lerly pointe out y proviing immeite feek on the ifferent oorinte views s the user interts (R2). Builing the Hierrhy Tree As esrie in this setion, the whole mehnism of reting n justing the segments relies on the iret mnipultion priniple (R3) to support quik n effetive wy of rilling own into multiple lotions of the tsets (R3). To enle quik lerning of the interfe, ll the opertions for eiting the hierrhy re performe on the min intertive view, y simple mouse intertion n one moifier key. A new segment is rete y iretly rushing (i.e. ynmi seletion in the intertive view) the ROI on ny existing series in the hierrhy (R3,). The user is require to right lik on the esire strting point new segment ppers on the ring lote one level outsie from the querie t piee, n the ROI is olore with light gry. The light gry highlight elimiting the ROI n its ssoite segment re ynmilly juste while the user rgs the ursor, until the mouse utton is relese t the intene ening point, finlizing the retion. When the user nees to fous on speifi time frme, she n perform time rushing on olletion of segments elonging to ring t one. Simultneous retion of s mny hil sequenes s segments present on ring n e one in single opertion y holing the Shift key while rushing the time frme on one of the ring s segments. Figure 3 epits how the multi-foi time intervls n e juste y iret mnipultion on the view (R3). A Ortho- Zoom mehnism [3] is use to just ROIs in the following wy: the user n look t erlier or lter t points in the prent series y rgging the ROI long the ring, keeping the sme time frme ut t ifferent time lotions (Fig. 3). Drgging the ursor on the ril iretion towrs the enter (Fig. 3) reues the intervl. Conversely, longer intervl is otine y rgging in the opposite iretion. Arrows ening the she urves tht link segment to its orresponing ROI re lso rggle n t s sliers to just the ROI size (Fig. 3). The r segment isply n lso e moifie: roun hnles ple on the extremities of eh 1740

5 CHI 2011 Session: Visul Anlytis f e Figure 3. Left: justing region of interest, () move n (,) resize. Right: justing t r segment, () hnge y-xis mplitue, (e) shif y-xis n (f) streth or squish time segment (Fig. 3f) llow for strething or squishing the t representtion, thus ffeting the time sle use for renering (the t lwys oupies ll the ville segment spe), the ROI n the t itself remin the sme. The user n lso set up speifi vlues y invoking the segment setting pop-up menu s she liks on the segment s time sle text lel on its top (Fig. 1g), oth the relte ROI n segment re upte in onsequene. Two more ontrollers (Fig. 3,e) expline lter, re provie to the user for t plot visul settings. Figure 4. Exmple of explortion of the NYSE exhnge ily open prie ( ) using the Preview moe: phntom of the mnipulte segment remins t the initil position (top left) s it is temporry overli on nother segment for tren omprison. The inner rings hve een shrunk to fous on the eepest levels of the hierrhy, n the reltionships etween the tsets re shown in the enter. Flexile Intertion for Visul Explortion Eiting the Hierrhy Tree Ajusting the Lyout for Segment Comprison Mintining the history of the explortion proess (R3f) is importnt to help users revisit importnt views n keep trk of their explortion [14]. Even though we o not expliitly support nvigtion history, the hierrhy tree mimis explortion history sine new segment is rete eh time the user wnts to rill own into the t n ll the existing segments remin in ple. When the user wnts to fous on the eepest levels of the hierrhy euse they orrespon to the finest etils, she oes not require the whole hierrhy tree to e isplye in etil. In the irulr representtion, the eepest levels re lote on the outer rings. To relese spe n fous visul ttention, the user n shrink the inner rings (see Fig. 4) y rgging ring towrs the enter of the irle view ring intertion (Fig. 5) is etile lter. The user n freely just the min intertive view through lssil pn (right utton rg), zoom (mouse wheel) n rotte (left utton rg) opertions to ess etils. When looking t time-series t, one of the min tsks is to ompre t t ifferent lotions in time. However, two segments n e spe out in the view, even showe upsie own euse of the irulr isply. This n mke the omprison more iffiult thn looking t sie-y-sie piees (R2). To filitte omprison, we give the user the full freeom of rerrnging the segments in the view: segment n e rotte nywhere in its own ring. Ring jumps re lso supporte. Thus the segments n e freely ligne either on stke onfigurtion y putting them on ifferent rings, or overli to otin overplotte t (Fig. 2) s they re semi-trnsprent. We isuss lter tht the overlppe re n e use for nlytil purposes y isplying the result of sttistil funtions inste of the rw t. To voi the view eing too rowe euse of the presene of too mny segments, the user n lso put t sie while keeping it t her isposl for lter. Doule liking on segment mkes it ollpse: segment of reue height, not isplying ny etils then tkes its ple. Hene visul lutter is reue, mking it possile to fous on smll suset of visully etile segments without losing the mentl mp of the urrent struture. The etils n e restore nytime y oule lik on ollpse segment. Figure 1 shows exmples of oth ollpse n etile segments. Keeping the whole history of explortion is sometimes not relevnt when intermeite stges re not importnt for the nlysis. For exmple when sequene of interest hs een ientifie, the user might no longer e intereste in its nestors in the hierrhy. To get ri of unuse segments, the user n enter the Deletion moe (isusse lter) n elete either single segment y liking on it, or segment n its whole su-hierrhy while holing the Shift key. The elete piees ispper in smooth fe out, voiing rupt hnges in the view when eleting whole rnh Looking for the unexpete is iffiult n teious n usully requires going through lot of opertions efore serenipitously isovering relevnt insights within the t. To voi the user messing up her lyout uring the Visul Explortion, we llow for Preview moe in whih the urrent lyout is mintine while the user performs temporry nlysis (e.g. rgging segment on top of nother). A phntom of the segment remins in ple t its initil position, n is restore s she releses the mouse utton. Figure 4 shows temporry omprison of similr trens of open prie t ifferent time sle n lotion y overlying one segment over the other in the Preview moe. KronoMiner is esigne to e s unonstrining s possile in terms of ompring multiple su-piees of severl tsets. Muh freeom is given to the user: segments from ny loe tset, t ny lotion in the tset n renere t ny time sle n e ple nywhere in the view. Esy ess

6 Segment Brnh Shift Ring All Selet 2 Collpse Exten Drg Drg Wheel Crete n just new segment Rotte Ring jump Rotte Shrink inner rings Pn Zoom Figure 5. Context-se intertion. The tivtion re is showe in pink, n the ffete elements in lue. Intertion n e performe on () single segment, () segment n its hilren rnh, () whole ring n () ll rings t the sme time. to the opertions through the iret mnipultion tehniques (R3) llows for quik n smooth mnipultion of the t with immeite visul feek. However, giving too muh freeom to the user lso hs its rwks. The hierrhy struture might not lwys e esy to re fter reorgnizing the segments. Visul ues hve to e integrte to help the user mintin her mentl mp. To this en, we use expliit she links, olor oing n immeite highlighting of the relte rnhing (nestors n hilren in the hierrhy) while hovering over segment. All these tehniques re esigne to give s mny visul ues s possile to keep wreness of the overll tree struture. Context-se Intertion We introue new ontext-se intertion lnguge for the user to e le to mnipulte more thn one segment t time. We exploit the urrent position of the ursor to etermine whether the user intene to perform n tion on ) single segment, ) segment n its whole rnh, ) single ring or ) the whole view. Figure 5 shows summry of the ifferent seletle groups, n the ville opertions for eh group using the mouse uttons n the Shift moifier key. The pink res orrespon to ll the ursor positions tht etermine group, whih is epite in lue. For exmple, y holing the Shift key n rgging segment, the user my performs rnh move opertion (Fig. 5) n then n rotte segment n ll its esennts in the sme intertion or mke them jump to ifferent rings. Context-se intertion llows the user to opt ifferent strtegies for Visul Explortion n optimize the intertion. For exmple, the rings n e use to tegorize the t; eh ring eing eite to speifi group the user wnts to keep relte (R3). By performing tions in the ring-moe, the user n pply the sme hnges to ll the elements of the group in single intertion while keeping onsisteny within the group. Keeping onsisteny while esigning iret mnipultion is possile lthough limite: there is finite numer of wys of ining opertions to the omintion of mouse uttons, mouse wheel n moifier keys. Other intertion tehniques suh s ontext menu or hotox [21] must e integrte s omplement to iret mnipultion to offer more omplex opertions n pilities. Following the sme ontext-se priniple just esrie (Fig. 5), we mke ville itionl funtionlity (isusse lter) on ontext menu tht is invoke y hitting spe r (Fig. 1e). The user n perform eletion opertions n visul settings tht ffet trgete set of segments, epening on the position where the menu hs een invoke. The menu lso ontins permnent options, suh s the ifferent moes ville, mong whih the Preview moe n the Deletion moe we introue erlier. The Detile Comprison Winow Cirulr lyouts suffer from istorting the t n might e onstrining for etile nlysis. To lne the rwks of the ril view, we hve esigne eite Detile Comprison Winow, lote on the right of the interfe (Fig. 1) where four selete segments n e isplye t time, in tritionl liner stke lyout. To isply segment in the omprison winow, the user is require to lik on the esire trgete spe mong the four ville, then lik on the esire segment to e isplye. If segment is lrey plotte, the new seletion reples it. The user n lso right lik on ny of the four liner plots to ler it. The Detile Comprison Winow ws initilly esigne s two sie-y-sie plots. The urrent tive plot ws permnently showe, the seon slot eing reserve for the segment uner the mouse ursor. One of our experts involve in the itertive esign proess si he [wnte] to e le to look t more thn two plots t time n tht they shoul e onstntly visile fter [he h] selete them. He lso reporte tht seeing them isplye on top of eh others woul filitte omprison s I nee them to e ligne. We urrently support four plots, ut one n esily imgine more to e visulize y mking srollle the etile view. To filitte t rowsing, the omprison winow is enrihe y intertion pilities: the user n pn the t in the liner plot for further explortion or to ontrol time lignment. The orresponing ROI n its ssoite segment re upte in the mentime. Two ontrollers (similr to Fig. 3,e) re ville on the uplite liner segments, to set up y-xis mplitue n shift. We gurntee full oorintion etween the two views so tht the user n esily keep trk of the impt of one opertion on the other view. Moreover, oth the r segment n its opy re highlighte while hovering over one or the other to mke it ler whih t is ommon to the oorinte views (R2). Dynmi Anlytil Methos When nlyzing time-series t, users re ommonly intereste in the evolution of their t over time for eteting trens n ptterns, ut lso nomlies. Anlytil methos n help gin insight into t y utomtilly eteting perioiity, reurrent ptterns n self-similrity [2]. However, one of the mjor prolems is tht most of these methos re hevily omin-epenent n thus nnot e pplie to ny type of t, n even less for ompring time-series from ifferent pplition omins. Moreover, the prmeters re often not self-explntory n hene not esy to set, espeilly without some prior knowlege of the tsets n the hrteristis of ptterns of interest. 1742

7 When the user hs no or vgue ie of the ptterns she is looking for, preise nlytil omputtions nswering speifi questions re of little use. In this se, the user hs to rely on Visul Explortion in orer to fin out the questions she oes not know yet. Bsi, generl n omin-inepenent nlytil omputtions (e.g. ifferene plot) re then ruil s they provie lues n suggestions of regions worth exploring eeper. Tools suh s TimeSerher2 [5] tht llow for query y exmple pttern fining, hve lrey emonstrte the enefit of integrting intertive visuliztion n nlytil pilities in single tool. KronoMiner enrihes its Visul Explortion pilities y integrting nlytil nlysis support in three ifferent wys: (1) visuliztion of the reltionships within n etween the tsets, (2) the Mgi- Anlytis Lens: on-the-fly omputtion n isply of nlytil methos ompring two segments, n (3) the Best Mth moe: on-the-fly omputtion n isply of the est mth of two segments. We etil them in the next setions. Showing the Reltionships Although lot of existing intertive visuliztions llow for pttern mining, to our knowlege, they ll only fous on single pttern t time. As onsequene, they fil in proviing the user with n overview of ll the reltionships existing within n etween the tsets. In other pplition omins, suh s Biology, it is ommon to represent the entire set of links representing reltionships (e.g. MizBee [23] isplys mthing susequenes of two hromosomes). Showing ll the reltionships t one help the user gin overll ie of oth the losely relte regions n isolte t, tht re usully strting points for explortion. In KronoMiner, we ompute similr ptterns using simple rute-fore similrity mining lgorithm t the loing of the tsets. Any other lgorithm or even preefine links n e use. Similrly to MizBee [23], the reltionships within n etween the ifferent entire tsets re isplye t the enter of the irulr lyout, onneting the t items of the min ring. The links intensity is mppe to the mthing sore (the rker, the higher). To mke trens visully ovious we omine the ege unling tehnique esrie in [16] with lph lening resulting from overlpping semitrnsprent links. In this wy, the user hs onstnt ess to reltionships informtion she n use for visully mining regions of interest (R4), therey guiing Visul Explortion. MgiAnlytis Lens Detile pirwise omprison is one of the min tsks when nlyzing time-series t. It usully involves omputing the orreltion etween the two series to ompre (or susequenes of them), sometimes t ifferent time sle; n viewing the similrity sores of isovere ptterns. Tritionl sttistil tools, suh s Mtl, llow for fine nlysis of the t: wie vriety of preefine nlytil methos re ville, s well s the possiility of efining new funtions. However, in suh tools the user is require to speify the ext instrutions (e.g. the ouns of the intervl to fous on) through the use of ommn line efore visulizing the result. In the se tht she oes not know extly wht to look for, she might generte lot of Figure 6. Exmple of use of the MgiAnlytis Lens on peer-to-peer tivity logs. A segment tht roughly orrespons to one y of tivity is use s lens. The lens isplys the orreltion plot, n the kgroun olor onveys the glol root men squres opertor on yellow-to-lue sle, in this se showing tht ily pttern is repete s the user rgs the lens long the series. plots efore fining something relevnt. Not only is the proess repetitive, hphzr n time-onsuming, it is lso istrting s the user splits her ttention fousing k n forth etween the ommn interfe n the visuliztions, therey requiring she rememers wht hs lrey een explore. Moreover, these tools usully require some si progrmming skills, whih nrrows the trget uiene. To voi reking the flow of the explortion proess n filitte opportunisti isovery, we introue the MgiAnlytis Lens, n intertive visuliztion tehnique tht omputes in rel time the result of funtion involving two timeseries, n immeitely isplys its result (R4). MgiAnlytis Lens is inspire y Mgi Lens [4] tht ffets the imge insie the ouns of shpe efining the lens. In ontrst to Mgi Lens tht uses eite winow s lens pplying fixe preefine opertor, MgiAnlytis Lens onsiers ll the t piees s potentil lens, s the isplye result epens on oth the t points in the winow use s lens, n the t points uner it. In Figure 6 the lens shows the ross-orreltion plot etween two series using the t ontine in the overlppe intervl. The plot is utomtilly upte while hnging the lignment s the user moves or strethes the segment use s lens. The kgroun olor of the lens is mppe to ompute glol mesure (root men squre in Fig. 6). This wy, the user n very quikly ssess if the urrent lignment ompres relte series. Any segment in the view n e use s MgiAnlytis Lens, mking it possile to ompre t oth within the originl series n etween the ifferent tsets, n lso t t ifferent time sles s the user n streth n squish the segments. A MgiAnlytis Lens n lso e plotte in the synhronize Detile Comprison Winow to ess the etils s showe in Fig. 1. The MgiAnlytis Lens is me ville through the ontextul menu (Fig. 1g). When entering the moe, ll the overli res re reple y the result of the selete funtions. The urrent prototype integrtes generl funtions suh s the ifferene plot, the minimum n mximum plots, n the inner prout or the root men squres for the kgroun, ut ny opertor n e emee. Ielly, the user shoul e le to efine her own funtions n then ustomize the tool y introuing new nlysis methos on emn (R4). 1743

8 CHI 2011 Session: Visul Anlytis e f Figure 8. The ifferent ville plot styles: () mplitue plot, () line hrt, () stter plot, () histogrm, (e) lines plot n (f) ule plot. Figure 7. Exmple of Best Mth etween two segments of NYSE exhnge ily open prie ( ) minimizing the root mens squre. An rh onveys the est mth lotion, its kgroun olor is mppe to the mthing sores on yellow-to-lue sle. The pre-ompute links re replite etween the two sequenes. In this exmple, the relshionships etween the segments surprisingly o not reflet the est mth lotion n might e worth exploring further. Fining the Best Mth There is n unnt literture out time-series sttistil tools tht re eite to pttern retrievl. Pttern isovery is n importnt tsk while nlyzing time-series t s it is the preliminry step to isover new wys of efining wht oul hrterize similrity. Allowing users to selet pttern from the t itself hs een showe to e useful [5], espeilly in the ontext of Visul Explortion when the user oes not hve preise ie of wht she is looking for. KronoMiner integrtes Best Mth moe (R4): the urrent selete segment efines the query pttern n hovering over segment efines it s eing the trget. The est mth oring to selete similrity mesure is ynmilly ompute n isplye: n rh links the query segment to its est mth within the trgete segment n the rest of the min view is fe out to voi visul lutter (Fig. 7). The kgroun olor of the rh shpe is mppe to the mthing sore to provie quik inition of the similrity. We lso replite the sme pre-ompute reltionships s showe in the enter of the ril isply. We thus mke ville the ontext tht woul otherwise e lost euse of fing out. The user n ess the etils nytime y hovering over the rh if she nees to know the preise sore vlue n etile informtion out the two mthe segments. Similrly to the MgiAnlytis Lens, the Best Mth tkes into ount the urrent time sle of the segments, thus llowing for ompring sequenes t ifferent time sle. For exmple, frtl ptterns within the sme time series oul e revele if the user etets high similrity etween segment n the series it omes from. The similrity funtion to e use is seletle through the ontext menu (Fig. 1g), ut ny other pttern mining metho oul e integrte. Plot Visul Settings For etile nlysis, the user is le to hnge the plot style mong the six ifferent ville (Fig. 8) in the ontext menu (Fig. 1g) other types suh s horizon grphs [13] or hetmps oul esily e e. Note tht the mplitue n ule plots prtiulrly suit for ril lyout s their pereption is less ffete y the irulr eformtion. We lso support t isply justments y llowing the user to lter the y-xis mplitue n shift. To o so, two rggle hn les re me ville on eh sie of segment in oth the irulr min view n the omprison winow (Fig. 3,e). Drgging up n own the left hnle respetively inreses n ereses the mplitue. The right hnle etermines the position of the origin of the y xis. Drgging it up n own shifts the hrt isply. All the renering settings in one of the two views re utomtilly reflete in the other, gurntying oorintion n onsisteny (R2). Figure 9. Exmple of use of the MgiAnlytis Lens on multivrite tset. Top plot: monthly verge ost of night s ommotion in Vitori, ottom: Consumer Prie Inex in Melourne ( ). The orreltion plot n the kgroun olor (root mens squre) inite ommon trens etween the two segments. One might hve to el with multiple vriles while keeping them ouple. Exmples inlue oorintes of moving ojet, min/mx tempertures n opening/losing prie of stok mrket tht woul not mke sense esynhronizing. We support multivrite tsets y piling up multiple views s showe in Figure 9. In this exmple, the MgiAnlytis Lens revels repete pttern, ut the tren is oserve t ifferent time sle. A room prie n the CPI evolve in similr wy t ifferent times, ut twie s fst the seon time. Experts oul mke the hypothesis tht there exists orreltion, n then potentilly preit the phenomenom when the pttern ours gin. Although it is possile to mke rings lrger to fit more thn 3 vriles t time, KronoMiner oes not sle to lrger imensions euse of visul lutter. Design improvements (e.g. ynmi reorering n stking) re neee to support high imensionl t. Spe-effiient lternte visuliztions (e.g. rie grphs [18]) woul prtilly solve the prolem ut imply issues for the nlytis lens esign. USER FEEDBACK The KronoMiner interfe hs een itertively refine through 4 prtiiptory esign sessions with 7 experts from ifferent omins t ifferent stges of the evelopment. Eh session fouse respetively on Computer Network (1 expert), Meil Siene (1), Wether Forest (3) n Finnes (2) n lste 4 to 5 hours uring whih we introue the tool, let the users try it n gthere feek n requirements.

9 Moreover, n expert user not involve in the esign proess hs teste the lst prototype on his own t in 2 working sessions of 3 hours eh. The prtiipnt is reserher stuying lrge sle peer-to-peer trffi t on vieo-on-emn wesites, n usully uses Mtl. After esriing the tool interfe n funtionlities, the user ws ske to explore his t using the think-lou protool. We first oserve the user ws lerly more omfortle with the interfe uring the seon interview, wheres he ompline it is iffiult to rememer, I hve to lern when testing the erlier prototype. At the en of the sessions, he ompre the lerning proess: It took me weeks, even months to lern how to use Mtl euse I h first to lern oe. I n o lot with your tool, n I on t even hve to think out it. We strive to mke the intertion s iret s possile so s not to rek the explortion flow. Context-se intertion requires trining, ut one nnot enefit from omplex tions while lwys keeping strightforwr intertions. We elieve novie user to e le to quikly get fmilir with the ifferent funtionlities ut this remins to e vlite in further stuies. Our expert s tset onsiste of month of the visiting popultion of two ifferent hnnels, smple from the server every 10 minutes. He first wnte to test the hypothesis tht the su-sequenes of everyy s t re similr within n etween hnnels. To o so, he rushe ROI on one hnnel orresponing to wht he esrie s the first y, roughly (he ws le to ientify it euse of the hrteristi shpe of the plot). To refine his seletion, he isplye the segment in the omprison winow, hnge the plot type to gin etter view n rely juste the time frme, pointing tht it oesn t hve to fit extly, I just wnt quik look. The user pnne the plot in the omprison winow to get n overll ie of y-to-y t n possily spot nomlies or interesting ehviour he wouln t see in the ril representtion. He juste the sle n the shift of the y-xis s the mplitue of the plot erese. At some point, he mentione the plot is too smll now, I n t sy if I see pttern, or if it is noise. So fr I ve een le to rely ientify similr pttern, ut now it s not possile nymore, n inresing the mplitue wouln t help. He thus eie to swith to the MgiAnlytil Lens moe s support to nlyze the ross orreltion. As he rgge the first y segment long the sequene, he oserve onstnt pek on the enter of the ompute plot (Fig. 6), initing high orreltion t zero ely. This justifie prt of his hypothesis n he efine the lens rgging s eing very hny. The prtiipnt roughly selete the lst y t points n rgge oth ys segments to outer rings to gin etter view, n me them overlp, still in the MgiAnlytis Lens moe. He then rgge the first segment to sequentilly ompre the t from the ltest y to the rest of the sequene, n vlite his hypotheses s they revele eing orrelte. While oing so, our user emphsize tht In Mtl you hve to generte tons of grphs. Even if you hve enough spe to isply them, you on t hve enough rin. Here, ll you hve to o is to simply rg [the segment]. The user strte to explore the t on the seonry hnnel. He foun the ily ross orreltion to e high etween the two tsets, ut not s strong s within the sme series. After extensive nvigtion n igging into the t, the work spe turne to e messy. Our user strte to rerrnge the segments: I nee to orgnize, shrunk inner rings, n ollpse some remining segments, s they were not of interest t this time I on t nee this guy for now.... When lening his spe, he relize he looke t mny plots without notiing: This sves me lot of time. Normlly I hve to eyell ll the t first euse I on t wnt to go into every piees euse it is teious. Further, he gthere unexpete finings s he ynmilly juste the rnge of the ROI to time intervls less thn y - The t roun minight is similr to tht roun noon! It s interesting. I woul hve expete it is the se t other rush hours, ut people shoul sleep t minight. It is worthwhile to stuy eeper. Our user ws enthusisti out nlytil pilities within tool supporting iret mnipultion. He si he wnte to explore more tsets with more vriles n longer time perios using our prototype, s it woul e too time-onsuming with tritionl tools. He mentione tht for nlyzing server t, he woul require to freely streth the time s servers likely hve ifferent hrwre. Hene the performne plots hve to e juste for fir omprison. The Best Mth Moe woul lso e helpful s the nee for mining guine inreses with the numer n size of the time-series. DISCUSSION Look: sexy or serious? When they o not propose soer interfe tht exhiits severl menus, ontrol pnels n numerous settings, tools for nlysis re not tken seriously s they o not look professionl. It is often tht these tools re mitte to e nielooking n plesnt to ply with, ut their pperne hs negtive impt on the users reline. A tool tht feels too esy to use might suffer from its pprent simpliity. Proxilly, users re highly emning for more intertivity n more user-frienly interfes n thus rete tension etween pprent profesionlism n essiility. KronoMiner emonstrtes tht it remins possile to perform effetive nlysis while offering high intertivity in semless tool. There is strong nee to onvine users to ept tht the sene of menus n settings pnels oes not men tool is not powerful. Our user who ws ffle t the eginning I ve never seen tht efore, ene up sking us for relese of the prototype to use it for his reserh. Bk to the Roots The previous oservtions rise the question of reking hits. Integrting fmilir intertion, n fmilir representtions seems to e essentil for the users to feel onfient n onsier the tool potentilly useful. Surprisingly, our expert hs not een le to strt ny onstrutive explortion uring our first meeting s he [ouln t] work without seeing somewhere some plots stke in liner view. However, he i not use the Detile Comprison Winow s muh s expete uring the seon interview. To his own surprise, he mitte tht it ws ressuring to e le to get 1745

10 k to the tritionl representtion, I know it is there, so I feel omfortle using the tool now. It is iffiult to mke n interfe feel fmilir sine people hve ifferent working hits, espeilly when trgeting wie uiene. We strive to esign KronoMiner to e s flexile s possile for multipurpose nlysis through prtiiptory esign stuies eliertely involving reserhers from isprte omins, s gurntee not to speilize to omin n n ttempt to over ommon nees. By oing so, we think our tool is pte to wie uiene. In return, it is still too generl to fully support eep nlysis, s our user quikly misse nlytil opertions he is use to pply - It woul e gret if you n more opertors to the lens. Allowing users to ustomize the tool hs revele to e ruil for funtionlities extension purpose. We lso elieve tht this woul help them ept more esily new wys of interting, s they oul personlize the tool to fit with their nees. CONCLUSION AND FUTURE WORK This pper introues KronoMiner, multipurpose tool for exploring time-series tsets. KronoMiner integrtes new intertion n visuliztion tehniques in semless tool esigne to e s unonstrining s possile so s to ese the explortion of multiple su-piees of interest for finer nlysis. Although it oes not solve ll the prolems, KronoMiner illustrtes how multi-foi nvigtion enrihe with generl sttistil tools n help users perform explortory nlysis on multiple time-series t. Feek from experts onfirme our elief tht users enefit from systems omining oth improve lgorithms n intertive visul interfes to ientify trens or spot nomlies. In prtiulr, the MgiAnlytis Lens tht provies n immeite visul feek hs revele to e promising tehnique. Despite the positive erly feek, longituinl stuies still hve to e run to vlite our lims n etter spot the limittions. We pln to provie ll our experts with relese of the tool for them to try it eeper on their own t n gther their feek. We will lso enompss mehnism for the user to e le to ustomize the nlytil tools to stuy how fr generl tool n support tsks the users re use to perform with their tritionl speilize systems. REFERENCES 1. W. Aigner, S. Miksh, W. Müller, H. Shumnn, n C. Tominski. Visulizing time-oriente t- systemti view. Comput. Grph., 31(3): , June W. Aigner, S. Miksh, W. Müller, H. Shumnn, n C. Tominski. Visul methos for nlyzing time-oriente t. IEEE TVCG, 14(1):47 60, C. Appert n J.-D. Fekete. Orthozoom sroller: 1 multi-sle nvigtion. In Pro. CHI 06, 21 30, E. A. Bier, M. C. Stone, K. Pier, W. Buxton, n T. D. DeRose. Toolglss n mgi lenses: the see-through interfe. In SIGGRAPH 93, 73 80, P. Buono, A. Aris, C. Plisnt, A. Khell, n B. Shneiermn. Intertive pttern serh in time series. In Pro. of Visuliztion n Dt Anlysis, , W. Chung, H. Chen, L. G. Choy, C. D. O Toole, n H. Atkhsh. Evluting event visuliztion: usility stuy of oplink sptio-temporl visulizer. Int. J. Hum.-Comput. Stu., 62(1): , P. Drgievi n S. Huot. Spirlok: ontinuous n non-intrusive isply for upoming events. In Pro. CHI 02, , G. M. Drper, Y. Livnt, n R. F. Riesenfel. A survey of ril methos for informtion visuliztion. IEEE TVCG, 15(5): , N. Elmqvist, Y. Rihe, N. H. Rihe, n J.-D. Fekete. Melnge: Spe foling for visul explortion. IEEE TVCG, 16(3): , J. Fils, A. Krlson, L. Shhmt, n B. Shneiermn. A visul interfe for multivrite temporl t: Fining ptterns of events ross multiple histories. Symposium On Visul Anlytis Siene An Tehnology, 0: , P. Fitts n J. Peterson. Informtion pity of isrete motor responses. Jrnl of Experimentl Psyhology, 67: , J. Go, Y. Co, W.-w. Tung, n J. Hu. Multisle Anlysis of Complex Time Series: Integrtion of Chos n Rnom Frtl Theory, n Beyon. Wiley-Intersiene, J. Heer, N. Kong, n M. Agrwl. Sizing the horizon: The effets of hrt size n lyering on the grphil pereption of time series visuliztions. In Pro. CHI 09, J. Heer, J. Mkinly, C. Stolte, n M. Agrwl. Grphil histories for visuliztion: Supporting nlysis, ommunition, n evlution. IEEE TVCG, 14: , H. Hohheiser n B. Shneiermn. Dynmi query tools for time series t sets: timeox wigets for intertive explortion. Informtion Visuliztion, 3(1):1 18, D. Holten. Hierrhil ege unles: Visuliztion of jeny reltions in hierrhil t. IEEE Trnstions on Visuliztion n Computer Grphis, 12: , W. Jve n N. Elmqvist. Stk zooming for multi-fous intertion in time-series t visuliztion. In PifiVis 10, 33 40, W. Jve, B. MDonnel, n N. Elmqvist. Grphil pereption of multiple time series. IEEE TVCG, 16(6): , E. J. Keogh n M. J. Pzzni. Relevne feek retrievl of time series t. In SIGIR 99, , R. Lopez-hernnez, D. Guilmine, M. J. Mguffin, n L. Brfor. A lyer-oriente interfe for visulizing time-series t from osillosopes. In PifiVis 10, 41 48, M. J. MGuffin n I. Jurisi. Intertion tehniques for seleting n mnipulting sugrphs in network visuliztions. IEEE TVCG, 15(6): , P. MLhln, T. Munzner, E. Koutsofios, n S. North. Liver: intertive visul explortion of system mngement time-series t. In Pro. CHI 08, , M. D. Meyer, T. Munzner, n H. Pfister. Mizee: A multisle synteny rowser. IEEE TVCG, 15(6): , W. Mller n H. Shumnn. Visuliztion methos for time-epennt t-n overview. In Pro. of the 2003 Winter Simultion Conferene, , C. Plisnt, B. Milsh, A. Rose, S. Wioff, n B. Shneiermn. Lifelines: visulizing personl histories. In Pro. CHI 96, , H. A. Rihter, J. A. Brotherton, G. Aow, n K. N. Truong. A multi-sle timeline slier for strem visuliztion n ontrol. Tehnil report, Georgi Institute of Tehnology, B. Shneiermn. Diret mnipultion: A step eyon progrmming lnguges. Computer, 16:57 69, B. Shneiermn. The eyes hve it: A tsk y t type txonomy for informtion visuliztions. In VL, , S. F. Silv n T. Ctri. Visuliztion of liner time-oriente t: A survey. Pro. of the Int. Conf. on We Informtion Systems Engineering, 310, C. Tominski, J. Aello, n H. Shumnn. Axes-se visuliztions with ril lyouts. In SAC 04: Pro. of the Symp. on Applie omputing, , J. W. Tukey. Explortory Dt Anlysis. Aison-Wesley, J. J. Vn Wijk. Cluster n lenr se visuliztion of time series t. In Pro. INFOVIS 99, 4 9, M. Weer, M. Alex, n W. Müller. Visulizing time-series on spirls. In Pro. INFOVIS 01, 7,