Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng BUSINESS PROCESS PERFORMANCE MANAGEMENT USING BAYESIAN BELIEF NETWORK Yeong-bn Mn 1, Yongwoo Shn 2, Km Jeehong 1, Dongsoo Km 3, Suk-ho Kang 1 1 Department of Industral Engneerng, Seoul, Natonal Unversty, +82-2-880-7360, bn01@snu.ac.kr; shkang@snu.ac.kr 2 Entrue Consultng BU, LG CNS, +82-2-880-7360, yongwoo.shn@lgcns.com 3 Department of Industral and Informaton Systems Engneerng, Soongsl Unversty, +82-2-820-0688, dskm@ssu.ac.kr Abstract As busness envronment changes dynamcally, effcent management strategy and rsk management are requred for a company to survve. To cope wth such requrements, BPMS (busness process management system) has been developed. BPMS enables companes to manage and mprove ther processes contnuously. Most of earler studes on busness process management have been focused upon process modelng, executon, and montorng. Therefore, there are a few researches that nvestgate how to mprove busness processes. Ths paper proposes a method for busness process performance management that ranges from busness actvty montorng to escalaton. To consder dependences between tasks, Bayesan belef network s employed and mathematcal model s desgned to determne the tasks to be escalated. Keywords: busness process management system (BPMS), key performance ndcator (KPI), Bayesan belef network (BBN), escalaton 1. Introducton As busness envronment changes dynamcally and competton becomes ferce, t s mportant for enterprses to handle rsks and to buld effcent management strateges. Under these changes n busness envronment, enterprses need to defne ther own crtcal success factors and key performance ndcators (KPIs) to evaluate the present state of operatons, and then they try to fnd the method for mprovng performance. Several performance measurement systems are n use today, and each has ts own group of supporters. One of the most nfluental approaches that have been mplemented n many companes s performance measurement whch s commonly based on the Balanced Scorecard (BSC) [7]. However, the performance measurement systems based on the BSC usually take a vertcal vew. Ths means that the structure of the BSC often mrrors the organzaton charts of enterprse to be measured. The ncluson of the organzatonal unts s mportant, but t s nsuffcent. Küng et al [9] proposed a performance measurement system consderng busness processes to take a horzontal flow. One of the most wdely used approaches consderng the horzontal flow s busness process management system (BPMS). BPMS extends the functonalty of workflow management systems (WfMS) beyond automaton nto areas such as analyss, montorng and cross-organzatonal nteractons [15]. The BPMS enables all stakeholders to have an understandng of an organzaton and ts performance, and to facltate process mprovement. The research ssues of the BPMS are categorzed n the perspectve of the process lfe cycle nto four groups: modelng, executon, montorng, and mprovement. Most of earler researches on the BPMS have been focused upon process modelng, executon, montorng. Therefore, there are a few researches that nvestgate how to mprove busness processes usng montored process data. And the researches related to process mprovement suggest general gudelnes through establshment of a framework rather than specfc escalaton methods. Also, most of them assume that tasks or components of a process are mutually ndependent. Ths paper proposes a method for busness process performance management that ranges from busness actvty montorng to mprovement. To consder dependences between tasks, Bayesan belef network s employed and mathematcal model s desgned usng key performance ndcators to 635
Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng evaluate performance of a task. The rest of the paper s organzed as follows. In secton 2, we revewed related work. Secton 3, present the busness process escalaton framework, and secton 4 descrbes the process escalaton strateges wth expermental results. Fnally, conclusons and future work are summarzed n secton 5. 2. Related work 2.1. Process montorng Montorng encompasses the trackng of ndvdual process nstances, so that nformaton on ther state can be vsualzed, and statstcs on the performance of processes can be provded. Ths nformaton can be used to work wth every partcpant n processes, so that problems n operatons can be dentfed and corrected. The degree of montorng depends on what nformaton the busness wants to evaluate and analyze and how busness wants t to be montored. Busness actvty montorng (BAM) coned by Gartner group s soluton for montorng of busness actvtes. BAM extendng montorng tools generally provded by BPMS refers to the aggregaton, analyss, and presentaton of real-tme nformaton about actvtes nsde organzatons. Most general approach for performance measurement usng operatonal data s to compare 'as-s' state and 'to-be' state. Based on the comparson of these two states, Roznat et al [14] suggested the method of workflow smulaton for operatonal decson support. Also, decson support method was proposed to buld rule-based event processng whch s consderng the dependences of state nformaton [8]. The exstng rule-based approaches defne the rules as a form of If condton Then acton. The rules are usually extracted from hstorcal log data or desgned by doman experts manually. To extract the meanngful correlatons, other methods whch are applyng decson tree [4], and genetc algorthm [1] were developed for montorng of the process nstances. However, these rule-based approaches have a lmt to evaluate the performance reactvely, rather than proactvely [6]. Rules are usually extracted from the attrbutes of completed nstances from hstorcal log data. However, an ongong nstance has only partal nformaton, composed of collected attrbutes of events untl observaton perod. Ths causes that a rule-based montorng system wats untl all the condtons of predefned rules are observed n ongong processes. 2.2. Process escalaton Accordng to the defnton of the WfMC process escalaton s a procedure executed when predefned constrants or condtons are not fulflled [16]. Ths means that the escalaton s an addtonal actvty to prevent montored process nstances from causng uncontrollable states of a process. As the abnormal states and ther escalaton methods can be dfferently defned accordng to the contexts of busness processes and envronments, t s dffcult to provde general soluton for the process escalaton. General approach to abnormal state analyss and escalaton method decson s smulatonbased. Smulaton s repeatedly performed to confrm whether a process nstance can be termnated successfully n varyng the value of the performance ndcators under the current condtons [11, 14]. After ths nvestgaton of the process nstance through smulaton, the target of escalaton s dentfed. And then, approprate actons wll be served to the nstance accordng to the extent of abnormalty on the performance ndcators. For example, ntensty of the escalaton can be controlled by the extent of overtme [2, 5, 12]. To determne the extent of the abnormalty, Grgor et al [3] set up the rule usng classfcaton technques lke decson tree and showed the escalaton taken by re-arrangement of the task prortes. In ths paper, we propose a method for busness process performance management that ranges from busness actvty montorng to mprovement. To evaluate the performance and abnormalty of tasks, cost and tme ndcators are consdered. And escalaton s conducted n a way that addtonal resources (e.g. cost and tme) are suppled to tasks. 636
Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng 3. Busness Process Escalaton Framework Fgure 1 llustrates the escalaton framework n BPMS. The abnormal state analyss and the escalaton engne are covered area n ths paper. Fgure 1. Escalaton framework n BPMS 3.1. Abnormal state analyss KPIs seen as a means of quantfyng the effcency and effectveness of tasks can be used to the current state of a process. As a process s a collecton of related, structured tasks that produce specfc actons, the evaluaton of a process can be done by aggregaton of task-level KPIs. However, t s unsutable to use task-level KPIs drectly to evaluate the performance of a process [3]. Because each task has dfferent KPIs and t s dffcult to dentfy relatonshps between the values of task-level KPIs n each task. Hence, each task s evaluated as a status correspondng to the combnaton of ts KPI values. In table 1, we can summarze the status of a task usng the followng states. state Best Good Normal Bad Worst Table 1. Status of a task descrpton All the KPI values n a task obtan good results The KPI values n a task lead postve effects to followng tasks General results when a task s ended Not serous, but the task concluded contans possble rsks affectng to followng tasks Montored results n a task leadng negatve effects to followng tasks Based on the hstorcal data of process executon, the status of a present task s determned. To formulate the status of a task, ordnary least squares lnear multple regresson s used. Table 2 descrbes the regresson model used n ths paper to determne the status of a task. At the end of every sngle task, the status of a task s evaluated assgnng KPI values to correspondng varables n the model. Table 2. Regresson model Y state value of th task jth KPI s normalzed value xj Y 0 2 x1 2 x1... n xn of th task j coeffcent 0~a a~b b~c c~d d~1 n number of KPIs n th task State Best Worst Bad Normal Good error term * a, b, c, d : cut-off value 637
Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng State predcton module whch to target tasks remaned or undone s operated every tme the status of a task s determned. It s assumed that a currently fnshed task or the determned status of a task can have nfluence on other unfnshed tasks. For example, f the status of a currently fnshed task s evaluated as worst, followng tasks are affected by the negatve status of the prevous task, so that t s lkely that these unfnshed tasks wll have lower performances. To consder these dependences between tasks, a Bayesan network s employed. The reason why we use a Bayesan network s as follows. As a Bayesan network can be represented as a drected acyclc graph, t s easy to model a process whch conssts of a sequence of connected tasks. A Bayesan network provdes a clear semantc nterpretaton of the model parameters. Unlke neural network models, whch usually appear to the user as a black box, all the parameters n a Bayesan network has an understandable semantc nterpretaton. It s easy to handle doman expert knowledge. In Bayesan modelng, doman knowledge can be coded as pror dstrbutons, pror meanng that probablty dstrbutons are defned before and ndependently of processng any possble sample data. Ths allows for combnng expert knowledge wth statstcal data n a very practcal way. From the doman knowledge and the hstorcal data, parameters needed to buld a Bayesan network are estmated. Usng ths network, the expected status of unfnshed tasks can be calculated probablstcally and we can nference last task n a gven process whether ends up successfully or not. 3.2. Escalaton To execute a task, tangble or ntangble resources related to busness process are requred. The knds of such resources are cost, tme, human resources, materals, etc. As avalable resources have a strong nfluence on the performance of a task, we assume that the more resources are assgned to a task, the better performance can be acheved. In ths paper, we only consder cost and tme. Because not only earler work done by [2, 10, 13] consdered cost and tme, but also these two resources are related to the performance ndcators drectly and ndrectly. Escalaton s a procedure whch s nvoked f last task n a process have a hgh probablty that wll be a negatve state. Escalaton s done by comparson between the worst state probablty of last task n a process and allowable lmt. In other words, based on the fnshed task, the worst state probablty of last task s calculated and then, escalaton s amed to lower ths probablty than maxmum lmt of the worst state probablty. Determnaton of tasks to be escalated s performed by followng model n table 3. Usng the model descrbed above, tasks to be escalated and necessary escalaton cost are determned. 4. Experments Experments are done to apply and evaluate the proposed escalaton method to the process composed of sequentally connected ffteen tasks. To show the effcency of the proposed escalaton method, escalaton polces are desgned as below. Rule-based escalaton Rule-based escalaton represented as a form of If condton Then acton s a most wdely used approach n BPMS. We consdered followng two rules. (1) If the prevous task s evaluated as bad or worst, escalaton wll be done to the followng task by rasng the normal state. (2) If the prevous two tasks are evaluated as (worst, worst) or (bad, worst), escalaton wll be done to the followng task by rasng the good state. 638
Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng N s x c jk w,, t jk w Table 3. Escalaton model a set of unfnshed tasks state of task decson varable whch ndcates whether unfnshed task wll be escalated or not addtonal cost and tme related to mprove the state of task from state j to k. The ndces j, k are on a scale of 1 to 5. The ndces are larger, the better performance can be acheved. w 1 weghts correspondng to escalaton cost and tme, where c wt c t Pworst ( sn s1, s2,..., sn 1) the worst state probablty of task n, derved from the Bayesan network P maxmum lmt of the worst state probablty B, T avalable budget and tme for escalaton mn w c x w t x (1) (1) Objectve functon s formulated as jk jk c t N j j k c 15 N j j k t 15 Pworst ( sn s1, s2,..., sn 1) P (2) N j j k N j j k c x B jk t x T jk x {0,1}, s { worst, bad, normal, good, best} (3) (4) a weghted sum of normalzed escalaton cost and tme. And t wll mnmze the effort for the escalaton. (2) the worst state probablty of last task n a process s less than allowable lmt (3) budget constrant (4) tme constrant No escalaton There are no escalaton actons to tasks n a process. Cost used n ths case s zero. Varyng the maxmum lmt of the worst state probablty, escalaton cost and the rato of nstances not fnshed n worst state are compared. Each escalaton polcy s executed 500 tmes. Results are shown at fgure 2. Fgure 2. Comparson experments As shown n fgure 2, the proposed escalaton method use less cost compared to other polces, because t has lttle chance of escalaton to tasks whch have low dependences wth others. Also, the proposed escalaton method shows hgh performance n the case of rgorous lmt. 639
Proceedngs of the 41st Internatonal Conference on Computers & Industral Engneerng 5. Conclusons In ths paper, we propose a method for busness process performance management that ranges from busness actvty montorng to escalaton. To consder dependences between tasks, Bayesan belef network s employed and mathematcal model s desgned to determne the tasks to be escalated. And we show that the proposed escalaton method s useful n the case of rgorous lmt and has a good performance compared wth the rule-based escalaton. There stll reman several further research ssues to be dealt wth. The model can represent varous structural features of process. In the model ntroduced here, tasks are connected sequentally. Processes n real world, however, can have other structures such as parallel, condton, and teraton. Also, technques to represent non-lnear relatonshps between performance ndcators can be consdered. Acknowledgement Ths work s supported by the Natonal Research Foundaton of Korea (NRF) grant funded by the Korea government (MEST) (No. 20110016160). References [1] M. Albaghdad, B. Brley, and M. Evens, "Event storm detecton and dentfcaton n communcaton systems," Relablty Engneerng & System Safety, vol. 91, pp. 602-613, 2006. [2] J. Eder, E. Panagos, and M. Rabnovch, "Tme Constrants n Workflow Systems," the Proceedngs of the 11th Internatonal Conference on Advanced Informaton Systems Engneerng, 1999. [3] D. Grgor, F. Casat, M. Castellanos, U. Dayal, M. Sayal, and M. C. Shan, "Busness process ntellgence," Computers n Industry, vol. 53, pp. 321-343, Apr 2004. [4] D. Grgor, F. Casat, U. Dayal, and M.-C. Shan, "Improvng Busness Process Qualty through Excepton Understandng, Predcton, and Preventon," the Proceedngs of the 27th Internatonal Conference on Very Large Data Bases, 2001. [5] E. Johann, P. Euthmos, P. Henz, and R. Mchael, "tme management n workflow systems," presented at the BIS'99, 1999. [6] B. Kang, S. K. Lee, Y.-b. Mn, S.-H. Kang, and N. W. Cho, "Real-tme Process Qualty Control for Busness Actvty Montorng," the Proceedngs of the 2009 Internatonal Conference on Computatonal Scence and Its Applcatons, 2009. [7] R.S. Kaplan, D.P. Norton, The Balanced Scorecard, Harvard Busness School Press, 1996. [8] A. Lundberg, "Leverage Complex Event Processng to Improve Operatonal Performance," Busness Intellgence, vol. 11, pp. 55-65, 2006. [9] P. Küng, A. Krahn, Buldng a Process Performance Measurement System some early experences, Journal of Scentfc and Industral Research, Vol. 58, No. 3/4, 1999. [10] R. S. Mans, N. C. Russell, W. van der Aalst, P. J. M. Bakker, and A. J. Moleman, "Smulaton to Analyze the Impact of a Schedule-aware Workflow Management System," Smulaton-Transactons of the Socety for Modelng and Smulaton Internatonal, vol. 86, pp. 519-541, 2010. [11] J. Nakatumba, A. Roznat, and N. Russell, "busness process smulaton: how to get t rght," ed: Sprnger-Verlag, 2010. [12] E. Panagos and M. Rabnovch, "Escalatons n workflow management systems," the Proceedngs of the workshop on on Databases: actve and real-tme, 1997. [13] O. O. Prsecaru, "Resource workflow nets: an approach to workflow modellng and analyss," Enterprse Informaton Systems, vol. 2, pp. 101-120, 2008. [14] A. Roznat, M. T. Wynn, W. M. P. van der Aalst, A. H. M. ter Hofstede, and C. J. Fdge, "Workflow smulaton for operatonal decson support," Data & Knowledge Engneerng, vol. 68, pp. 834-850, 2009. [15] W. M. P. van der Aalst, M. Rosemann, and M. Dumas, "Deadlne-based escalaton n process-aware nformaton systems," Decson Support Systems, vol. 43, pp. 492-511, 2007. [16] Workflow Management Coalton, Workflow Management Coalton Termnology & Glossary Doc.-Nr: WFMC-TC-1011, 1999. 640