Explanng Produc Release Plannng Resuls Usng Concep Analyss Gengshen Du, Thomas Zmmermann, Guenher Ruhe Deparmen of Compuer Scence, Unversy of Calgary 2500 Unversy Drve NW, Calgary, Albera T2N 1N4, Canada {dug, zmmerh, ruhe}@cpsc.ucalgary.ca Absrac Objecve: Ths paper ams o generae explanaons from a seres of daa pons obaned from a decson suppor sysem called ReleasePlanner for supporng produc release plannng and consdered o be a black box. Mehod: Concep analyss s appled o 1085 daa pons receved from runnng 10 scenaros of a real world produc release plannng projec wh 35 canddae soluons generaed by ReleasePlanner. Resuls: Three man resuls are obaned: (1 paerns beween npus and oupus; (2 mpac of ndvdual npu parameers on oupus; and (3 sensvy level of oupus n dependence of npus. Concluson: Concep analyss s shown o be a feasble echnque for ganng more nsgh no he srucure of resuls obaned from a black box npu-oupu sysem, such as, bu no lmed o, ReleasePlanner. Keywords Explanaons, concep analyss, produc release plannng 1 Inroducon Produc release plannng nvolves decson makng on assgnng feaures o dfferen releases for ncremenal sofware produc developmen. I mus smulaneously consder several aspecs, such as conflcng sakeholder prores and objecves, feaure nerdependences, and resource and rsk consrans [15]. A decson suppor sysem called ReleasePlanner [13] has been developed o suppor decson makers n he complex release plannng process. I s based on compuaonally effcen opmzaon algorhms for he generaon of a se of alernaves soluon havng a proven degree of opmaly. However, he fndngs from a seres of expermens conduced wh ReleasePlanner users revealed ha hey were relucan o accep he soluons advsed by hs ool [6]. Smlar observaon has also been made on oher sysems [4] [10]. I was concluded n [1] ha he major problems are no echncal problems, bu people problems n whch people have very lmed undersandng on he suppor hey ge from decson suppor sysems. In addon, accordng o [2], produc release plannng problem s classfed as a wcked problem [14] whch s hard o be precsely formulaed. Thus he procedure needed o solve produc release plannng problems (as demonsraed n ReleasePlanner s more complex and requres more n deph explanaons o acheve good user undersandng on he ool suppor and s soluons. How can we faclae beer undersandng of he ReleasePlanner soluons? In hs paper, a daa analyss echnque called concep analyss [11] s appled for hs purpose. I s appled o nvesgae daa whn a specfc produc release plannng problem and denfes hdden relaonshps beween he projec npus and oupus. In parcular, hree ypes of relaonshps are analyzed: Paerns beween he npu and oupu arbues Impac of ndvdual npu arbues on he oupus Sensvy level of he oupus o he npus The answers o hese hree research quesons provde addonal knowledge ha s currenly unavalable o users of he ReleasePlanner sysem. As a resul, he users accepance and rus level on he ool and s soluons s expeced o be mproved. The remander of hs paper s organzed as he follows. Secon 2 gves an overvew of produc release plannng and he relaed decson suppor ool ReleasePlanner. Secon 3 nroduces he background of concep analyss. In Secon 4, a sample produc release plannng projec s nvesgaed o demonsrae he applcaon of concep analyss. Secon 5 analyzes and nerpres he resuls n he conex of he hree saed quesons. Fnally, Secon 6 summarzes he research and oulnes fuure research. 2 Produc Release Plannng Many formal approaches have been proposed for produc release plannng, such as ncremenal fundng mehod [5], cos-value approach [9], plannng sofware evoluon wh rsk managemen [8], and hybrd nellgence (EVOLVE* [15]. The laer s used n hs paper. Ths secon gves a shor overvew of hs approach o he exen necessary o undersand and judge he resuls obaned from concep analyss presened laer n hs paper. More deals on he mehod are avalable from [15]. 2.1 Techncal Formulaon In ncremenal produc developmen, he goal of produc release plannng s o selec from a se of feaures F = {f 1,,f n } and o assgn hem o one of K possble releases each of hem havng a wegh (relave mporance of ξ k (k = 1 K. A release plan s descrbed by vecor x wh
decson varables {x(1,,x(n}, where x( = k f feaure f s assgned o release opon k {1,,K}; and x( = K+1 oherwse (.e. he feaure s posponed. Two ypes of feaure dependences are consdered: couplng and precedence relaonshp. A couplng CC(f, f j ndcaes ha boh feaures f and f j mus be released jonly. A precedence PC(f, f j ndcaes ha feaure f canno be released laer han f j. Some feaures can be fxed o ceran release by he pre-assgnmen preassgnx(=k, ndcang ha f s fxed o release k. The plannng approach consders T resource ypes for mplemenng he feaures. Capaces Cap(k, relae each release k o each resource ype {1,,T}. Every feaure f requres an amoun of resources of ype r(f,. Thus, each release plan x assgns feaure f o release k expressed as x( = k, for all releases k and resource ypes, mus sasfy Σ x(=k r(f, Cap(k,. A se of sakeholders S = {s 1,,s q } s nvolved n release plannng. Each of hem has a relave mporance λ {λ 1,,λ q }. I s a nne-pon ordnal scale ha provdes dfferenaon n he degree of mporance. The hgher he mporance value s, he more mporan he sakeholder s. In bref, he purpose of release plannng s o provde he mos aracve feaures a he earles releases o he mos mporan sakeholders. For he purpose of hs paper, Value(s, f, Urgency(s, f, and Compeveness(s, f are he hree arbues of a feaure s aracveness. Each feaure can be prorzed from hese hree crera wh he value rangng from 0 o 9. These crera are assocaed wh he weghs μ 1, μ 2, and μ 3, respecvely. The hree prorzaon crera are he bass o formulae he objecve of produc release plannng. The objecve funcon Uly(x s defned as a lnear combnaon of he prory voes of sakeholders relaed o hese crera: Uly( x = ( ξ K Pr ory( k f k= 1 : x( = k where Prory(f s an aggregaed prory of f defned as: S q Prory( f = ( λ s ( μ1 Value( s, f + s= s1 μ Urgency s, f + μ Compeveness( s, f 2 ( 3 2.2 ReleasePlanner ReleasePlanner [13] s a decson suppor sysem ha ams a performng sysemac produc release plannng based on compuaonally effcen opmzaon algorhms. Users are able o perform wha-f analyss o pro-acvely explore dfferen scenaros defned by a sequence of npus of he same projec under nvesgaon. In addon, he ool s capable of generang a se of dversfed soluon alernaves for each nsance. A seres of sudes on ReleasePlanner revealed ha s users ended o have hgher confdence and rus on her manual soluons han he ones generaed more effcenly by he ool [6]. The major reason s ha he ool works n a black box manner and he raonale of soluon generaon s hard o undersand by he users. Ths s even more complcaed because he users usually nvesgae mulple scenaros wh mulple soluons. 3 Concep Analyss Concep analyss, frsly nroduced n [17], s a heory of daa analyss o denfy concepual srucures among a se of daa. I has been successfully appled o many felds [12], ncludng n sofware engneerng [11]. In hs paper, concep analyss s nvesgaed o address he hree research quesons presened n Secon 1. Anoher wo echnques,.e. rough se analyss and dependency nework analyss, have also been appled o explan release plannng soluons by ReleasePlanner [7]. However, hey can only deal wh he frs wo research quesons and are no he focus of hs paper. Dealed applcaons of hese wo echnques are avalable a [7]. 3.1 Basc Termnology Concep analyss nvesgaes he relaons R beween a se of objecs O, and a se of arbues A. The rple C = (R, O, A s called a formal conex. Def. 1 (Common Arbues and Common Objecs: For any se of objecs b O, he se of common arbues havng he same arbue value s called common arbues relaed o b and s denoed by ca(b = {a A o b : (o, a R}. For a se of arbues A A, her common objecs are co(a = {o b a A : (o, a R}. Def. 2 (Formal Concep: Each par c = (b, A wh b = co(a and A = ca(b s called a formal concep. I demonsraes a paern,.e. relaon, beween b and A. Def. 3 (Concep Lace: All formal conceps for a gven conex C are called a complee concep lace n whch conceps can be parally ordered. If c 1 = (b 1, A 1 and c 2 = (b 2, A 2 are wo conceps n he conex C, a paral order c 1 c 2 can be defned whenever b 1 b 2 and A 1 A 2. Def. 4 (Greaes Lower Bound and Leas Upper Bound: The greaes lower bound of c 1 and c 2 s he concep wh objecs b 1 b 2 and arbues held by all objecs n b 1 b 2. The leas upper bound of c 1 and c 2 s he concep wh arbues A 1 A 2 and objecs whch have all arbues n A 1 A 2. 3.2 An Illusrae Example Appled o plannng produc releases, O consues he se of feaures F o be assgned o dfferen releases. The npu o and oupu from produc release plannng usng ReleasePlanner form se A. Fgure 1 shows a smple example of concep analyss n hs doman. The upper par s a daa able of feaure se F = {f 1,,f 4 } defned wh he arbue se A = {a 1,,a 3 }. In hs able, he value of each arbue for each feaure can be H, M, or L.
These values represen dfferen value ranges. The lower par of hs fgure shows he correspondng concep lace wh all he conceps {c 1,,c 8 } and her order relaons. In hs lace, he values of he arbues n each concep are also hghlghed. Among all he conceps, c 1 s he mos general one and c 8 s he mos specfc one. Some of he order relaons among he conceps are c 2 c 1, c 3 c 1, and c 7 c 1. From hs fgure, we can also denfy he leas upper bound and greaes lower bound of a se of conceps. For example, c 1 and c 4 are he leas upper bound and greaes lower bound of c 2 and c 3, respecvely. c 2 = ({f 1,f 2 },{a 1 :H, a 3 :L} c 4 = ({f 1 },{a 1 :H, a 2 :H, a 3 :L} Daa Table (H: Hgh M: Medum L: Low c 5 = ({f 2 },{a 1 :H, a 2 :M, a 3 :L} Concep Lace c 1 = ({f 1, f 2, f 3, f 4,},{a 3 : L} c 3 = ({f 1,f 3 },{a 2 :H, a 3 :L} c 6 = ({f 3 },{a 1 :L, a 2 :H, a 3 :L} Fgure 1: Example concep analyss c 7 = ({f 4 },{a 1 :M, a 2 :L, a 3 :L} c 8 =({Ǿ},{a 1 :H, a 1 :M, a 1 :L, a 2 :H, a 2 :M, a 2 :L, a 3 :H, a 4 :M, a 3 :L} 4 Applyng Concep Analyss o Explan Produc Release Plannng Resuls 4.1 Sample Projec To llusrae he applcaon of concep analyss o explan resuls generaed by ReleasePlanner, we nvesgae on a sample projec based on he daa from a real lfe produc release plannng problem. As a summary, hs projec s defned wh he followng npus: F = 31 feaures {f 1,, f 31 } o be assgned K = 2 releases S = 18 sakeholders {s 1,,s 18 } wh weghs {λ 1,, λ 18 } Prorzaon crera Urgency(s,f, Value(s,f, and Compeveness(s,f T = 3 ypes of resources {Res1, Res2, and Res3} The full deals of hs projec seng can be referred o hp://pages.cpsc.ucalgary.ca/~dug/concepanalyss. Ths projec seng and he resuls obaned from are aken as he baselne scenaro. From hs baselne, he ool user also generaes anoher 9 scenaros ha he user hnks o be he mos mporan (bu no necessarly he complee scenaros for nvesgaon. Togeher hese 10 scenaros are used for wha-f analyss whch s useful and mporan for produc release plannng, as dscussed n Secon 2.2. For all hese scenaros, n oal 35 soluons are generaed by ReleasePlanner for laer analyss. 4.2 Daa for Concep Analyss Wh he above projec sengs, each feaure f n each soluon s assocaed wh a daa pon used for concep analyss (see Table 1. The selecon of he arbues s based on he experence of manual analyss of several produc release plannng projecs [15]. The frs sx npu arbues are relevan o sakeholder voes whch are consdered n he objecve funcon for plannng. In parcular, ConfUrgency(f, ConfValue(f, and ConfComp(f are he sandard devaon beween dfferen sakeholders voes for each feaure f from he hree crera, respecvely. They ndcae he degree of dsagreemen among sakeholder opnons. The oher sx npu arbues address resource ulzaon and crcaly of feaures. Table 1: Daa defned for concep analyss Inpu Arbue Defnon AverageUrgency(f λs Urgency( s, f S = S1 AverageUrgency( f = AverageValue(f S18 λs S = S1 AverageComp(f (smlar for AverageValue(f, AverageComp(f RelConfUrgency(f ConfUrgency( f Re lconfurgency( f = RelConfValue(f AverageUrgency( f RelConfComp(f (smlar for RelConfValue(f, RelConfComp(f r( f,re s Res URao(f Re surao( f = 31 ( = 1, 2, 3 r( f,re s Res Crcaly(f ( = 1, 2, 3 S18 = 1 Re s Crcaly( f = Re s URao( f 2 k = 1 Oupu Arbue Defnon Release(f Release(f r( f, Re s Cap( k,re s Based on our prevous experence on he analyss of hese arbues, each arbue s dscrezed accordng o Table 2. The purpose of dscrezaon s o scale arbues wh connuous values o a nomnal or ordnal scales. Table 2: Dscrezaon of he defned arbues Inpu Arbue Value Range Dscrezaon AverageUrgency(f Hgh [6, 9] AverageValue(f Medum [4, 6 n [0, 9] AverageComp(f Low [0, 4 RelConfUrgency(f Hgh [0.7, 1.0] RelConfValue(f Medum [0.4, 0.7 RelConfComp(f Res URao(f Res Crcaly(f n [0, 1] n [0, 1] n [-1, 1] Low [0.0, 0.4 Hgh [0.10, 1.00] Medum [0.05, 0.10 Low [0.00, 0.05 No [0.00, 1.00] Low (-0.01, 0.00 Medum [-1.00, -0.01] Oupu Arbue Value Range Dscrezaon Release(f An neger n [1, 3] No necessary
Based on he above defnon and dscrezaon, a able wh 1085 daa pons (35 soluons wh each conanng 31 feaures s obaned for laer concep analyss. The complee able s avalable a he webse provded earler. 4.3 Concep Analyss of he Daa We used an open source lbrary called Colbr/Java [3] o perform concep analyss. I bulds a concep lace whch conans all paerns (conceps for he daa prepared n Secon 4.2. We hen mplemened a ool o raverse he concep lace o selec only hose paerns where he dsrbuon of he oupu values sgnfcanly changed along he (subse relaons beween hese paerns. To es sgnfcance, we used Fsher Exac Value and Ch Square ess (sgnfcance level of p=0.01 [16]. Usng our ool, wo flered laces were generaed ha conan he paerns whch affec he dsrbuon of releases he mos: Concep lace #1 conans 64 paerns where he lkelhood of assgnng a feaure f o release 1 s ncreased by a leas 45%. Concep lace #2 conans 1093 paerns where he lkelhood of assgnng a feaure f o any release,.e. 1, 2 or 3 (posponed, s ncreased or decreased by a leas 30%. The frs lace s essenally a par of he second one. The deals of hese laces are avalable a he webse gven earler and wll be analyzed n deph n Secon 5. Table 3: Example record n he generaed concep laces Conex Res3Crcaly(f _Low Var AverageComp(f _Hgh Δ_R1 0.49 Conex_R1 548 Conex+Var_R1 118 Δ_R2-0.3 Conex_R2 339 Conex+Var_R2 1 Δ_R3-0.18 Conex_R3 198 Conex+Var_R3 0 Each flered lace consss of a number of paerns and ransons beween hese paerns n he form shown n Table 3. Ths able s read as, for all he 1085 cases n he daase, he dsrbuon of feaures f followng he paern of Res3Crcaly(f = Low ( conex par s 548 daa pons for release 1 ( conex _R1 ; 339 for release 2 ( conex_r2 ; and 198 for release 3 ( conex_r3. The paern of Res3Crcaly(f = Low AND Average Comp(f = Hgh ( conex and var s suppored by 118 daa pons for release 1 ( conex+var_r1 ; 1 for release 2 ( conex+var_r2 ; and 0 for release 3 ( conex+var_r3. The ranson beween hese wo paerns can be undersood as a rule: addng Average Comp(f = Hgh ( var o he conex par ncreases he lkelhood of assgnng a feaure f o release 1 by 49% ( Δ_R1, and decreases he lkelhood o release 2 and 3 by 30% ( Δ_R2 and 18% ( Δ_R3, respecvely. 5 Analyss and Inerpreaon of Resuls In hs secon, we analyze he wo laces generaed n Secon 4.3 from hree perspecves: smlary of paerns, mporance of npu arbues, and sensvy of oupus. The fndngs from hese aspecs conan new knowledge ha reveals he underlyng relaonshps beween he npus o ReleasePlanner and s oupus, for he suded sample projec. They can be used as explanaons for hs decson suppor sysem and s soluons. 5.1 Paern Transons and Daa Smlary Each generaed concep lace covers he mos sgnfcan paerns dscovered from he produc release plannng daa used for concep analyss. These paerns are presened n he conex par and of he dfferen granulares,.e. from he mos general o he mos specfc. A general paern can be ransformed o more specfc ones, and vce versa. By examnng he generalzaon or specfcaon relaonshps among hese paerns, he ransons among he paerns become vsble. In addon, he dscovered paerns demonsrae he smlares shared among he daa used for analyss. Daa ha are caegorzed under a same paern are of he smlary as demonsraed by he paern. For any wo paerns ha can be generalzed o he same more general paern, he wo daa ses supporng hese paerns mus be smlar o each oher n he way ha s represened from he general paern. To llusrae he paern ranson and daa smlary n hs sample projec, he concep lace #1 s analyzed for smplcy. Any oher laces can be analyzed smlarly. Fgure 2 shows he op four levels of paerns whn hs concep lace and he ransons of hese paerns. The complee ransons of all he paerns n hs lace can be referred o he webse provded earler. In hs lace, he mos general paern s #1, as shown n he very op of he fgure. I can be specfed o paern #2, #3, and #4 a he second level. In hs case, we say paern #1 s he generalzaon of paern #2, #3, and #4. On he oher hand, paern #2, #3, and #4 are he hree specfcaons of paern #1. Each of hese hree paerns can be furher specfed o oher paerns unl no more specfc paern can be found. For example, one of he mos specfc paerns s paern #60. I follows he specfcaon pah of paern #1 #3 #6 #41 #50 #60. 7 {R1R,R2R, R1C,R3C} Legends: 2 {R1C,R3C} 3 {R1R,R3C} 4 {R2R,R3C} AU: AverageUrgency(f 5 {R1R,R1C,R3C} 6 {R1R,R2R,R3C} 8 {R3R,R1C, R2C,R3C} 1 {R3C} 16 {AC,R1R, R3R,R3C} AV: AverageValue(f R2R: Res2URao(f 17,18 {RCC,R3R, R1C,R3C} RCU: RelConfUrgency(f RCV: RelConfValue(f RCC: RelConfComp(f R1R: Res1URao(f R1C: Res1Crcaly(f AC: AverageComp(f R3R: Res3URao(f R2C: Res2Crcaly(f R3C: Res3Crcaly(f Fgure 2: Transon of paerns (concep lace #1
Regardng he smlares shared among all he 1085 daa used for he analyss, all hese daa are he same n erms of her values on he npu R3C (Res3Crcaly(f, as llusraed hrough he mos general paern,.e. paern #1. More smlary s dscovered from he daa #1 o #715 and #869 o #930 because hese daa pons also share he same value on he npu arbue R1C (Res1Crcaly(f, besdes on R3C. As a resul, hese daa form a more specfc paern,.e. paern #2. These resuls can be nerpreed as a ype of explanaons on ReleasePlanner soluons. If, n a soluon, he release assgnmen of a feaure s suppored by general paern(s ha are suppored by a large number of daa pons, he users are more lkely o accep such resul. Oherwse, hey mgh wan o furher mprove he soluon. 5.2 Imporance Level of Inpus on Oupus By examnng all he found paerns ( conex par, we can denfy each npu arbue s mporance level o he oupu arbue, n our case he release. The assumpon s ha he hgher he number of he occurrence of an npu n he paerns s, he more mporan hs npu s n deermnng he release value. However, an excepon o hs assumpon s ha hs number canno be as hgh as he oal number of daa used for analyss. The raonal for hs excepon s gven laer. For hs purpose, we nvesgae he second concep lace whch provdes more coverage han he frs one on he paerns nheren n he daa. Fgure 3 summarzes he number of occurrence of each npu n hs concep lace. Res3Crcaly(f appears o be he mos mporan arbue. I s n all he paerns and has he same value. In oher words, has no nfluence a all on he dsrbuon of release. Res1Crcaly(f and Res1URao(f are mporan arbues whch have dfferen values. The leas mporan ones are AverageComp(f, AverageUrgency(f, and RelConfComp(f. Oher npu arbues have medum level of mporance. 1200 1000 800 600 400 200 0 # of Occurrence of Each Inpu Arbue (n All Exsng Paerns In all paerns AverageUrgency(f AverageValue(f AverageComp(f RelConfUrgency(f RelConfValue(f RelConfComp(f Res1URao(f Res2URao(f Res3URao(f Res1Crcaly(f Res2Crcaly(f Res3Crcaly(f Fgure 3: Imporance level of each npu on he oupu (concep lace #2 Ths par of he resuls provdes he ReleasePlanner users wh he explanaons by denfyng a subse of all he defned npus ha play he mos sgnfcan mpacs on he ool when generaes soluons. 5.3 Sensvy Level of Oupus o Inpus The generaed concep laces also check f addng a new npu arbue ( var par o he exsng paerns ( conex par would change he dsrbuon of release. If he change s sgnfcan, hs npu s lkely responsble for such change,.e. he oupu s sensve o hs npu. To observe he sensvy of he oupu on each npu, he second concep lace s used for analyss agan. In parcular, we examne sx ypes of how he var par may mpac on he dsrbuon of releases: R1/R2/R3 +0.30: ncrease by a leas 30% n he dsrbuon of assgnng a feaure f o release 1, 2, and 3, respecvely R1/R2/R3-0.30: decrease by a leas 30% n he dsrbuon of assgnng a feaure f o release 1, 2, and 3, respecvely For each npu, we frs coun s number of occurrence n he var par of each record. For example, n he record n Table 3, he npu arbue AverageComp(f s n he var par wh 118 cases supporng he mpac of R1 +0.30. Therefore s number of occurrence n hs record, for hs ype of mpac, s 118. Then, by summng up such numbers for all he records of he same mpac ype, we oban he oal number of occurrence of hs npu. Fgure 4 shows hs number for each npu based on he above calculaon. We assume ha he hgher hs number s, he more sensve he oupu s o hs npu. 8000 7000 6000 5000 4000 3000 2000 1000 0 # of Occurrence of Each Inpu Arbue (as New Inpu Arbue o Exsng Paerns R1 +0.30 R1-0.30 R2 +0.30 R2-0.30 R3 +0.30 R3-0.30 AverageUrgency(f AverageValue(f AverageComp(f RelConfUrgency(f RelConfValue(f RelConfComp(f Res1URao(f Res2URao(f Res3URao(f Res1Crcaly(f Res2Crcaly(f Res3Crcaly(f Fgure 4: Sensvy level of he oupu o each npu (concep lace #2 From hs fgure, RelConfUrgency(f, Res3URao(f and AverageComp(f have he mos sgnfcan mpacs on he dsrbuon of release. Specfcally, release 1 s mos sensve o RelConfUrgency(f and AverageComp(f for a leas 30% of ncreased dsrbuon, and o AverageComp(f for a leas 30% of
decreased dsrbuon. Release 2 s mos sensve o Res3URao(f and AverageComp(f for a leas 30% of ncreased dsrbuon. Bu hey usually have no mpac as R2-0.30 or R3 ±0.30. On he oher hand, Res1URao(f, Res1Crcaly(f, and Res3Crcaly(f almos never conrbue o any change by a leas ±30% n any release. Alhough hey occur he hghes mes n he paerns n Fgure 3, her sensvy levels are no as sgnfcan as a leas ±30% and canno be refleced n Fgure 4. The oher npu arbues n general have medum level of sensvy on he oupu arbue. The above resuls explan some sensvy aspecs of he soluons generaed by ReleasePlanner. Ths knd of knowledge reveals he degree of mpac from changng dfferen npu parameers. In case of unceran daa, he rule of humb s ha he more robus a soluon, he hgher chance of accepance by he user. 6 Conclusons and Fuure Work In hs paper, a formal daa analyss mehod called concep analyss s combned wh sascal hypohess esng o explan complex soluons recommended by ReleasePlanner, a decson suppor sysem for produc release plannng. The resuls of our analyss of he daa of ndvdual release plannng projecs conan addonal knowledge ha s currenly unavalable o he ool users. Specfcally, such knowledge explans he underlyng relaonshps nheren n he nvesgaed daa, n erms of he underlyng paerns beween he npu and oupu daa, as well as he mporance and sensvy levels of npus on oupus. These explanaons nend o acheve beer undersandng on he soluons of ReleasePlanner, and herefore hgher accepance level from he user sde. To demonsrae he applcaon of concep analyss and sascal hypohess esng, a sample produc release plannng projec was nvesgaed. Alhough he fndngs presened n hs paper are specfc o he sample projec, he mehodology of applyng such analyss s generc (snce reas he decson suppor sysem as a black box and can be appled o any oher produc release plannng projecs, or oher sofware sysems n whch explanng complex soluons o users s necessary. As a very mporan fuure work, emprcal sudes wll be conduced wh ReleasePlanner users n order o jusfy he usefulness and effecveness of he proposed mehod for explanng he ool s soluons. In addon, he resuls obaned from he concep analyss, as presened n hs paper, only provdes one ype of explanaons on ReleasePlanner soluons and s by no means complee. The explanaons generaed from hs mehod are beer o be used wh oher ypes of explanaons (e.g. he ones dscussed n [7] ha address he soluons from dfferen aspecs. Therefore we wll also nvesgae on how hese dfferen ypes of explanaons obaned from dfferen echnques are complmenary o each oher so ha hey can ogeher provde he ool users wh a more complee vew of explanaons on he ool and s soluons. Acknowledgemen The auhors would lke o hank he Naural Scences and Engneerng Research Councl of Canada (NSERC and he Albera Informacs Crcle of Research Excellence (CORE for her fnancal suppor of hs research. Many hanks are due o Danel Gözmann and Chrsan Lndg who provded he Colbr/Java mplemenaon. References [1] M. J. A. Berry, G. S. Lnoff: Daa Mnng Technques: For Markeng, Sales, and Cusomer Relaonshp Managemen (2 nd Edon, Wley, 2004. [2] P. Carlshamre: Release Plannng n Marke-Drven Sofware Produc Developmen: Provokng an Undersandng. Journal of Requremens Engneerng, Vol. 7, 2002, pp. 139-151. [3] Colbr/Java, avalable a hp://code.google.com/p/ colbr-java/, las accessed February 2008. [4] F. Davs, J. Koemann: Deermnans of Decson Rule Use n a Producon Plannng Task. Journal of Organzaonal Behavor and Human Decson Processes, Vol. 63, No. 2, 1995 pp. 145-157. [5] M. Denne, J. Cleland-Huang: The Incremenal Fundng Mehod: Daa Drven Sofware Developmen. IEEE Sofware, Vol. 21, No. 3, 2004, pp. 39-47. [6] G. Du, J. McElroy, G. Ruhe: A Famly of Emprcal Sudes o Compare Informal and Opmzaon-based Plannng of Sofware Releases. Proceedngs of he 5 h Inernaonal Symposum on Emprcal Sofware Engneerng, Ro de Janero, Brazl, 2006, pp. 212-221. [7] G. Du, G. Ruhe: Comparson of Two Machne Learnng Technques for Explanng Resuls n Produc Release Plannng. Submed o Journal of Informaon Scences, Specal Issue on Applcaons of Compuaonal Inellgence and Machne Learnng o Sofware Engneerng, 2008, 36 pages. [8] D. Greer: Decson Suppor for Plannng Sofware Evoluon wh Rsk Managemen. Proceedngs of he 16 h Inernaonal Conference on Sofware Engneerng and Knowledge Engneerng, Banff, Canada, 2004, pp. 503-508. [9] J. Karlsson, K. Ryan: A Cos-Value Approach for Prorzng Requremens. IEEE Sofware, Vol. 14, No. 5, 1997, pp. 67-74. [10] L. Lehola, M. Kauppnen, S. Kujala: Requremens Prorzaon Challenges n Pracce. Proceedngs of he 4 h Inernaonal Conference on Produc Focused Sofware Process Improvemen, Vol. 3009, 2004, pp. 497-508. [11] C. Lndg, G. Snelng: Assessng Modular Srucure of Legacy Code Based on Mahemacal Concep Analyss. Proceedngs of he 19 h Inernaonal Conference on Sofware Engneerng, Boson, USA, 1997, pp. 349-359. [12] U. Prss: Formal Concep Analyss n Informaon Scence. Annual Revew of Informaon Scence and Technology, Vol. 40, 2006, pp. 521-543. [13] ReleasePlanner, avalable a www.releaseplanner.com, las accessed Aprl 2008. [14] H. W. J. Rel, M. M. Webber.: Dlemmas n a General Theory of Plannng. Polcy Scences, Vol. 4, 1973, pp. 155-169. [15] G. Ruhe, A. Ngo-The: Hybrd Inellgence n Sofware Release Plannng. Journal of Hybrd Inellgen Sysems, Vol. 1, No. 2, 2004, pp. 99-110. [16] L. Wasserman: All of Sascs: A Concse Course n Sascal Inference (2 nd Edon, Sprnger, 2004. [17] R. Wlle: Resrucurng Lace Theory: an Approach based on Herarches of Conceps. In: Ordered Ses (Ed. I. Rval, Redel, Dordech-Boson, 1982, pp. 445-470.