Expert Systems with Applications 36(2009)8938-8945 Contents lists available at Science Direct Expert Systems with Applications ELSEVIER journalhomepagewww.elsevier.com/locate/eswa Recommender system for software project planning one application of revised cbr algorithm Heng-Li Yang a Cheng-Su Wang b -Department of Management Information Systems, National Cheng-Chi University, 64 Section 2, Chihnan Road, Mucha Dist, Taipei 116, Taiwan Department of Information Management, Chung Hua University, 707, Sec. 2, WuFu Rd, Hsinchu, Taiwan ARTICLE INFO ABSTRACT Project management is an experience-driven and knowledge-centralized activity Therefore, project man- gers require some assistance to reduce the uncertainty at the early stage of constructing project plans. To overcome the predicament faced by project managers, this investigation proposes a hierarchical cri- Recommender system Multiple objective decisions thermore, to solve HCa problems, a revised case-based reasoning(RCBR)algorithm, is presented and recommender system for software project planning is implemented, based on multiple objectives deci- sion techniques and the mining approach. Finally, the proposed RCBr algorithm is successfully applied nalyze 41 re from a software consultancy in Taiwan. Experimental results demonstrate that RCBR can efficien rs to construct project plans at an arly stage. Additi the knowledge discovery process of RCBr provides project managers with Its similar to what-if analysis. The knowledge can enable project managers to obtain feasible infor- mation to re-schedule project resources, and bargain with their customers in the early project planning e 2008 Elsevier Ltd. All rights reserved. 1 Introduction and related work Raymond and Bergeron(2008)found that project management information system(PMIS) directly influences project success by Information technology(rr) or information system(IS)project complying with project deadlines and technical specifications. The management is regarded as a knowledge-centralized and experi- PM certainly requires additional assistance even when not applying ence-driven activity that is hard for inexperienced staff to imple PMIS (Raymond Bergeron, 2008). Garcia, Roman, Penalva, and ent(Lee Lee, 2006). The Standish Group CHAOS report nilla(2008)addressed this problem using association mining indicated that only 47% of software project could deliver within rules that determine meaningful project strategies from discrete budget and required functions(Standish Group International, numerical project data. Additionally Stamelos and Angelis(2001) 004). Additionally, Tesch, Kloppenbborg, and Frolick(2007)even successfully applied analogy-based methodology for cost estimation und that only 28% of It projects are successfully completed on across multiple software projects by project portfolios. The above time. IT project management is a difficult and challenging job, of- researchers intended to improve information transparency and ten with a low success rate. Therefore, project managers(PMs)of quality via prediction(such as budget or project duration prediction) It/is projects indeed require some assistance to do their jobs well and IT systems(such as PMis or knowledge sharing mechanism). (Dweiri Kablan, 2006). However, the entire IT project manage- Conversely, many researchers have focused on project risk analysis, ment process is highly uncertain and full of risks, and success rate which has become the primary research stream within project man- of a project generally depends on various critical factors. A system- agement. Wallace, Keil, and Rai(2004)and Han and Huang(2007 atic methodology is required to solve problems faced by PMs. derived software project risk factors by cluster analysis and empiri- Some researchers have, in recent decades, addressed project cal surveys, respectively. Both these two researches found that nent issues from different perspectives. Some researchers requirement and project planning and control are closely related cused on improving project information transparency and to project requirement risk and complexity risk level (from low to quality to enable PMs to increase the project success rate. high). Moreover, Rodriguez-Repiso, Setchi, and Salmeron(2007 onsidered these factors as critical success factors, and adopted fuz- zy cognitive map to build an It project success model. all the cited E-mail addresses: yanhenccuedu tw(.-L Yang), swang@chu.edutw(C-s. investigations above involve predicting the level of project risk, and then evaluating the project success rate. 0957-4174 front matter o 2008 Elsevier Ltd. All rights reserved. do:101016/ Jeswa200811.050
Recommender system for software project planning one application of revised CBR algorithm Heng-Li Yang a,*, Cheng-Su Wang b aDepartment of Management Information Systems, National Cheng-Chi University, 64 Section 2, Chihnan Road, Mucha Dist., Taipei 116, Taiwan bDepartment of Information Management, Chung Hua University, 707, Sec. 2, WuFu Rd., Hsinchu, Taiwan article info Keywords: Project planning Case-based reasoning Hybrid AI applications Recommender system Multiple objective decisions abstract Project management is an experience-driven and knowledge-centralized activity. Therefore, project managers require some assistance to reduce the uncertainty at the early stage of constructing project plans. To overcome the predicament faced by project managers, this investigation proposes a hierarchical criteria architecture (HCA) to enable project managers to describe project requirements adequately. Furthermore, to solve HCA problems, a revised case-based reasoning (RCBR) algorithm, is presented and a recommender system for software project planning is implemented, based on multiple objectives decision techniques and the mining approach. Finally, the proposed RCBR algorithm is successfully applied to analyze 41 real projects from a software consultancy in Taiwan. Experimental results demonstrate that RCBR can efficiently provide related information to help project managers to construct project plans at an early stage. Additionally, the knowledge discovery process of RCBR provides project managers with results similar to what-if analysis. The knowledge can enable project managers to obtain feasible information to re-schedule project resources, and bargain with their customers in the early project planning stage. 2008 Elsevier Ltd. All rights reserved. 1. Introduction and related work Information technology (IT) or information system (IS) project management is regarded as a knowledge-centralized and experience-driven activity that is hard for inexperienced staff to implement (Lee & Lee, 2006). The Standish Group CHAOS report indicated that only 47% of software project could deliver within budget and required functions (Standish Group International, 2004). Additionally, Tesch, Kloppenbborg, and Frolick (2007) even found that only 28% of IT projects are successfully completed on time. IT project management is a difficult and challenging job, often with a low success rate. Therefore, project managers (PMs) of IT/IS projects indeed require some assistance to do their jobs well (Dweiri & Kablan, 2006). However, the entire IT project management process is highly uncertain and full of risks, and success rate of a project generally depends on various critical factors. A systematic methodology is required to solve problems faced by PMs. Some researchers have, in recent decades, addressed project management issues from different perspectives. Some researchers have focused on improving project information transparency and project quality to enable PMs to increase the project success rate. Raymond and Bergeron (2008) found that project management information system (PMIS) directly influences project success by complying with project deadlines and technical specifications. The PM certainly requires additional assistance even when not applying PMIS (Raymond & Bergeron, 2008). Garcia, Roman, Penalvo, and Bonilla (2008) addressed this problem using association mining rules that determine meaningful project strategies from discrete numerical project data. Additionally, Stamelos and Angelis (2001) successfully applied analogy-basedmethodology for cost estimation across multiple software projects by project portfolios. The above researchers intended to improve information transparency and quality via prediction (such as budget or project duration prediction) and IT systems (such as PMIS or knowledge sharing mechanism). Conversely, many researchers have focused on project risk analysis, which has become the primary research stream within project management. Wallace, Keil, and Rai (2004) and Han and Huang (2007) derived software project risk factors by cluster analysis and empirical surveys, respectively. Both these two researches found that requirement and project planning and control are closely related to project requirement risk and complexity risk level (from low to high). Moreover, Rodriguez-Repiso, Setchi, and Salmeron (2007) considered these factors as critical success factors, and adopted fuzzy cognitive map to build an IT project success model. All the cited investigations above involve predicting the level of project risk, and then evaluating the project success rate. 0957-4174/$ - see front matter 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2008.11.050 * Corresponding author. Tel.: +886 2 29387651; fax: +886 2 29393754. E-mail addresses: yanh@nccu.edu.tw (H.-L. Yang), cswang@chu.edu.tw (C.-S. Wang). Expert Systems with Applications 36 (2009) 8938–8945 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
H.-L Yang C-S. Wang/Expert Systems with Applications 36(2009)8938-8945 A software project management has four stages, namely project construction project. They retrieved previous cases from a historical planning, scheduling, monitoring and control. To evaluate the pro- integrated project database, then stored them into a knowledge base ject success, both the project risk factor and project information for quality designing The retrieved solutions from past cases may quality issue should be addressed in all four stages. Each stage, par- need to be revised for further application. The successful problem- ticularly project planning, is closely related to project success(Lee solving experiences are then retained for further reuse. Traditional lee, 2006). a project without adequate project planning has a CBR algorithms have 4R-reasoning steps(Aamodt Plaza, 1994). igh risk level (Han Huang, 2007: Wallace et al., 2004). Thus, a However, although the project recommender system may provide comprehensive project planning could lead to project success. the relevant case(s) for PMs referencing. the revised directions of However, project planning should consider different project attri- these retrieved It projects remain ambiguous. PMs still do not know butes, such as type size and customer requirements. The situation how to adjust these projects to generate a reasonable project plan of project planning is highly uncertain and vague during the early prototype. stage. Therefore, PMs need a lot of assistance and information his study presents a revised CBR mechanism, named RCBR, Tina, Ma, and Liu(2002) integrated knowledge rules into a that integrates the traditional CBR algorithm with mining ap- mathematical model for R&D to assist project selection for Nation proaches. RCBR consists of two stages. In stage I, RCBR retrieves Science Foundation of China. However, mathematical models typ- the fittest case(s) for planning reference according to the query cally make too many assumptions to be applied well. Since PM equirement of PMs In stage Il, RCBR mines the cases retrieved equently construct project plans based on past experiences, Lee in stage I to provide further refined and potential knowledge. the and Lee(2006) recommended adopting case-based reasoning recommendation result of RCBr could decrease the uncertainty (CBR) algorithm to provide previous project cases requiring the in early planning stages, and thus assist PMs to plan projects suc least modification for constructing new project plans. CBR appears cessfully. Additionally, this investigation has the following goals: a feasible algorithm to solve the pi multiple conflicting objectives. Unfortunately, project (1) to enable PMs to consider their project requirements care- requirements of real applications are too complicated to describe lly, so that the retrieved case(s) fit reality: via multiple objectives. A sophisticated model is required to permit (2)to revise the traditional CBR approach, and apply rCBr to PMs to present their project requirements adequately Addition reduce uncertainty in project planning, and ly, PMs applying the proposed model of Lee and Lee (2006)might 3)to provide some clues(or information) for revised direction, ave no idea how to revise the retrieved case. even if it is the fitte thus improving the usefulness of the online application. one. Yang, Yin, Lin, and Pan(2007)claimed that only the actionable PMs should obtain refined information, rather than raw cases, from presents a new problem description called hierarchical criteria the case base. Therefore, this study develops a systematic method- architecture(HCA). Section 3 describes the proposed revised-CBR ology to assist project planning. algorithm. Section 4 summarizes the experimental results of the PMs should be aware of all attributes of projects, such as project proposed model. Section 5 presents conclusions and recommenda- size(Garcia, Quintales, Penalvo, Martin, 2004: Martin, Pearson, tions for future research. Furumo, 2007), cost (Love Irani, 2003), budget, customer require lents( Dweiri Kablan, 2006)and customer profiles(Dweiri& Ka- 2. Suggested problem presentation: hierarchical criteria blan, 2006) when constructing project plans. PMs should ideally perform what-if analysis to build the project plan prototypes, such as recovery plans and staff schedules. These activities are experi- The problem descriptions in traditional CBr algorithms have to follow the structure and stored format of case descriptions. In complicated and difficult due to globalization(Mahaney Lederer. other words, PMs have to raise their queries according to the case 2003). For instance, resources(such as staff, information and knowl- structure, regardless of their real requirements. For instance, Fig. 1 edge)might need to be transferred among countries and time zones. depicts a traditional CBR, CBR-Works, which formulates PMs'que- Therefore, early planning of iT projects is increasingly important. ies in a multi-objective format(Coello, 2000 ). However, as the Artificial intelligence(Al)approaches can effectively reduce the decision conditions become increasingly complex, the problem uncertainty, thus enabling PMs to handle the early stage of project descriptions in one-level multiple objectives become too rough to planning comfortably. Donzellli(2006)presented a decision sup- represent the decision problem correctly and completely. The solu port system(DSS) to assist PMs to predict and simulate proje tion is difficult to trust if the decision problem cannot be described implementation processes. Donzellli's DSs successfully improved well. the efficiency of project management and the schedule control Adomavicius and Tuzhilin (2005) stated that a next-generation quality. Dweiri and Kablan(2006)also applied fuzzy decision-mak- recommender system should be able to solve multiple dimensional ing techniques to project management. They proposed a fuzzy sys- problems, which are complex and close to the real situations faced tem to evaluate and predict the efficiency of project management by decision makers. To provide PMs with the most appropriate through project cost, project time and project quality. case(s) for project planning references, a new problem presenta- his study introduces an Al approach, called case-based reason- tion model is stipulated to permit PMs to focus adequately on their ing(CBr), to decrease uncertainty in early stages of project planning. project requirements. As is well-known, CBR is particularly useful for solving ill-defined This study proposes a novel problem description, named hierar and unstructured problems(Belecheanu. Pawar, Barson, Bredehorst, chical criteria architecture(HCA), to enable decision makers to de- Weber, 2003). CBR is the most appropriate Al technique for pro- scribe their decision problems adequately. As implied by the name, ducing project planning recommendations at early stages(Lee HCA allows decision makers to describe each decision objective Lee, 2006). since IT (or IS) projects are typically described in an with multiple levels to their desired level of details. If a problem unstructured manner, such as in the case-format. CBR is a paradigm, is described in HCA, then decision makers can drill each problem concept and instinctive mechanism for problem solving. Similar to dimension down until the required level of detail is reached. the human problem solving process, CBR retrieves past experiences Additionally, PMs could assign different relative weight for reuse on a target problem(Yang Wang, 2008). Cirovic and Cekic attribute to represent its significance. Fig. 2 illustrates PMs eval (2002)adopted CBR to support the preliminary design phase of a ating the case similarity in terms of three dimensions, " customer
A software project management has four stages, namely project planning, scheduling, monitoring and control. To evaluate the project success, both the project risk factor and project information quality issue should be addressed in all four stages. Each stage, particularly project planning, is closely related to project success (Lee & Lee, 2006). A project without adequate project planning has a high risk level (Han & Huang, 2007; Wallace et al., 2004). Thus, a comprehensive project planning could lead to project success. However, project planning should consider different project attributes, such as type, size and customer requirements. The situation of project planning is highly uncertain and vague during the early stage. Therefore, PMs need a lot of assistance and information. Tina, Ma, and Liu (2002) integrated knowledge rules into a mathematical model for R&D to assist project selection for Nation Science Foundation of China. However, mathematical models typically make too many assumptions to be applied well. Since PMs frequently construct project plans based on past experiences, Lee and Lee (2006) recommended adopting case-based reasoning (CBR) algorithm to provide previous project cases requiring the least modification for constructing new project plans. CBR appears to be a feasible algorithm to solve the project planning problem with multiple conflicting objectives. Unfortunately, project requirements of real applications are too complicated to describe via multiple objectives. A sophisticated model is required to permit PMs to present their project requirements adequately. Additionally, PMs applying the proposed model of Lee and Lee (2006) might have no idea how to revise the retrieved case, even if it is the fittest one. Yang, Yin, Lin, and Pan (2007) claimed that only the actionable information would provide reference value for decision makers. PMs should obtain refined information, rather than raw cases, from the case base. Therefore, this study develops a systematic methodology to assist project planning. PMs should be aware of all attributes of projects, such as project size (Garcia, Quintales, Penalvo, & Martin, 2004; Martin, Pearson, & Furumo, 2007), cost (Love & Irani, 2003), budget, customer requirements (Dweiri & Kablan, 2006) and customer profiles (Dweiri & Kablan, 2006) when constructing project plans. PMs should ideally perform what-if analysis to build the project plan prototypes, such as recovery plans and staff schedules. These activities are experience-driven and have uncertain results, and become increasingly complicated and difficult due to globalization (Mahaney & Lederer, 2003). For instance, resources (such as staff, information and knowledge) might need to be transferred among countries and time zones. Therefore, early planning of IT projects is increasingly important. Artificial intelligence (AI) approaches can effectively reduce the uncertainty, thus enabling PMs to handle the early stage of project planning comfortably. Donzellli (2006) presented a decision support system (DSS) to assist PMs to predict and simulate project implementation processes. Donzellli’s DSS successfully improved the efficiency of project management and the schedule control quality. Dweiri and Kablan (2006) also applied fuzzy decision-making techniques to project management. They proposed a fuzzy system to evaluate and predict the efficiency of project management through project cost, project time and project quality. This study introduces an AI approach, called case-based reasoning (CBR), to decrease uncertainty in early stages of project planning. As is well-known, CBR is particularly useful for solving ill-defined and unstructured problems (Belecheanu, Pawar, Barson, Bredehorst, & Weber, 2003). CBR is the most appropriate AI technique for producing project planning recommendations at early stages (Lee & Lee, 2006), since IT (or IS) projects are typically described in an unstructured manner, such as in the case-format. CBR is a paradigm, concept and instinctive mechanism for problem solving. Similar to the human problem solving process, CBR retrieves past experiences for reuse on a target problem (Yang &Wang, 2008). Cirovic and Cekic (2002) adopted CBR to support the preliminary design phase of a construction project. They retrieved previous cases from a historical integrated project database, then stored them into a knowledge base for quality designing. The retrieved solutions from past cases may need to be revised for further application. The successful problemsolving experiences are then retained for further reuse. Traditional CBR algorithms have 4R-reasoning steps (Aamodt & Plaza, 1994). However, although the project recommender system may provide the relevant case(s) for PMs referencing, the revised directions of these retrieved IT projects remain ambiguous. PMs still do not know how to adjust these projects to generate a reasonable project plan prototype. This study presents a revised CBR mechanism, named RCBR, that integrates the traditional CBR algorithm with mining approaches. RCBR consists of two stages. In stage I, RCBR retrieves the fittest case(s) for planning reference according to the query requirement of PMs. In stage II, RCBR mines the cases retrieved in stage I to provide further refined and potential knowledge. The recommendation result of RCBR could decrease the uncertainty in early planning stages, and thus assist PMs to plan projects successfully. Additionally, this investigation has the following goals: (1) to enable PMs to consider their project requirements carefully, so that the retrieved case(s) fit reality; (2) to revise the traditional CBR approach, and apply RCBR to reduce uncertainty in project planning, and (3) to provide some clues (or information) for revised direction, thus improving the usefulness of the online application. The remainder of this study is organized as follows. Section 2 presents a new problem description called hierarchical criteria architecture (HCA). Section 3 describes the proposed revised-CBR algorithm. Section 4 summarizes the experimental results of the proposed model. Section 5 presents conclusions and recommendations for future research. 2. Suggested problem presentation: hierarchical criteria architecture The problem descriptions in traditional CBR algorithms have to follow the structure and stored format of case descriptions. In other words, PMs have to raise their queries according to the case structure, regardless of their real requirements. For instance, Fig. 1 depicts a traditional CBR, CBR-Works, which formulates PMs’ queries in a multi-objective format (Coello, 2000). However, as the decision conditions become increasingly complex, the problem descriptions in one-level multiple objectives become too rough to represent the decision problem correctly and completely. The solution is difficult to trust if the decision problem cannot be described well. Adomavicius and Tuzhilin (2005) stated that a next-generation recommender system should be able to solve multiple dimensional problems, which are complex and close to the real situations faced by decision makers. To provide PMs with the most appropriate case(s) for project planning references, a new problem presentation model is stipulated to permit PMs to focus adequately on their project requirements. This study proposes a novel problem description, named hierarchical criteria architecture (HCA), to enable decision makers to describe their decision problems adequately. As implied by the name, HCA allows decision makers to describe each decision objective with multiple levels to their desired level of details. If a problem is described in HCA, then decision makers can drill each problem dimension down until the required level of detail is reached. Additionally, PMs could assign different relative weights to each attribute to represent its significance. Fig. 2 illustrates PMs evaluating the case similarity in terms of three dimensions, ‘‘customer”, H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945 8939
H.-L Yang C-S. Wang/ Expert Systems with Applications 36(2009)8938-8945 hoisin Fig. 1. Problem description via multiple objectives: CBR-Works interface. and"project keywords" For"customer capabilities" PMs Table 1 sure similarity in terms of separate"hardware"and"soft- Variable definition of our model imensions. Each dimension is given a different importance variable weight in order to describe complex requirements. In contrast, the Total number of case(s) within the case base traditional multi-objectives problem description does not permit The ith case within the case base, i=1, 2..n PMs to depict their problems at this level of detail. If the case is de eature (ie attribute) describing the scribed in HCA, then the queries can be handled hierarchically. The m retrieved case thus corresponds well to real problems, and pro- fety Target(T) that a PM wishes to query for recommendation vides additional reference value to pms The jth feature where 3. Proposed model - the revised CBR I fet identified jth feature of Target This study assumes that previous IT(or IS)projects have been wge.n et) he level of fey of fety Weight of feature j:j=1,2,-,m and T, i=1, 2,-n, derived by Eq (2) stored in a case base for future reference. The projects may be The gap of fet, between case and T, stored in the case base in an unstructured manner. and not in obtained by Eq. (1) any particular storage format. The proposed application procedure is as follows. First, PMs describe the project requirement according Fet_check _function The feature-check function, defining the similarity of fet, to reality via HCA. The HCa project requirement is then compared vith the pervious projects in the case base using the revised CBR (RCBR)algorithm. Finally, the most similar case(s)are retrieved identified through HCA. Therefore, a new evaluation process is nec- for reference. Furthermore, these retrieved cases are analyzed by essary to compare the case base with the target provided by the data mining techniques, e. g association analysis. The resulting PMs. Furthermore, the retrieved case(s)are sent directly to the analyzes provide refined actionable information to PMs. Table 1 PMs for reuse. Alternatively, PMs may revise the solution. The suc- lists the definitions of the variables in rcbr cess experience of project plan construction is then sent back to Fig 3 depicts the proposed two-stage revised CBR model. This the case base for further reuse. This is the scenario of well-known RCBR model follows the traditional CBR 4R(retrieve, reuse, revise CBR-4R cycle. However, some mechanisms should be altered to and retain)reasoning steps In stage 1, RCBR solves the problem deal with the HCA problem. Properties w1.2.1 Hardware Custome W31 Manufacture DSS Fig. 2. The project plan construction problem in HCA
‘‘project” and ‘‘project keywords”. For ‘‘customer capabilities”, PMs may measure similarity in terms of separate ‘‘hardware” and ‘‘software” dimensions. Each dimension is given a different importance weight in order to describe complex requirements. In contrast, the traditional multi-objectives problem description does not permit PMs to depict their problems at this level of detail. If the case is described in HCA, then the queries can be handled hierarchically. The retrieved case thus corresponds well to real problems, and provides additional reference value to PMs. 3. Proposed model – the revised CBR This study assumes that previous IT (or IS) projects have been stored in a case base for future reference. The projects may be stored in the case base in an unstructured manner, and not in any particular storage format. The proposed application procedure is as follows. First, PMs describe the project requirement according to reality via HCA. The HCA project requirement is then compared with the pervious projects in the case base using the revised CBR (RCBR) algorithm. Finally, the most similar case(s) are retrieved for reference. Furthermore, these retrieved cases are analyzed by data mining techniques, e.g., association analysis. The resulting analyzes provide refined actionable information to PMs. Table 1 lists the definitions of the variables in RCBR. Fig. 3 depicts the proposed two-stage revised CBR model. This RCBR model follows the traditional CBR 4R (retrieve, reuse, revise and retain) reasoning steps. In stage I, RCBR solves the problem identified through HCA. Therefore, a new evaluation process is necessary to compare the case base with the target provided by the PMs. Furthermore, the retrieved case(s) are sent directly to the PMs for reuse. Alternatively, PMs may revise the solution. The success experience of project plan construction is then sent back to the case base for further reuse. This is the scenario of well-known CBR-4R cycle. However, some mechanisms should be altered to deal with the HCA problem. Fig. 1. Problem description via multiple objectives: CBR-Works interface. Customer Project Project Keywords Budget Duration HR Accounting Manufacture DSS Hardware Software Target W1 W2 W3 w1,1 w1,2 w2,1 w2,2 w3,1 w3,2 w3,3 w3,4 w1,2, 1 w1,2, 2 Capability Properties Fig. 2. The project plan construction problem in HCA. Table 1 Variable definition of our model. Variable Definition and description n Total number of case(s) within the case base Ci The ith case within the case base, i = 1,2,...,n fet Feature (i.e., attribute) describing the case m Number of features describing each case in the case base T Target (T) that a PM wishes to query for recommendation fetj The jth feature where fetCi j identified jth feature of case fetT j identified jth feature of Target ( wgtfet Weight of feature j; j = 1,2,...,m Sim (Ci, T) Similarity between casei and T, i = 1,2,...,n, derived by Eq. (2) gap(fetCi j ; fetT j ) The gap of fetj between casei and T, i = 1,2,...,n, j = 1,2,...,m, obtained by Eq. (1) Level (fetj) The level of fetj Next_level (fetj) The next level of fetj Fet_check_function The feature-check function, defining the similarity of fetj 8940 H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945
H.-L Yang C-S. Wang/Expert Systems with Applications 36(2009)8938-8945 8941 Case base nput Target Reuse Revised Stage Il Reuse Fig 3. Proposed RCBR model Fig. 4 depicts a recursive algorithm(named fet_ckf_Rewrite portion distribution) enables decision makers to mechanism)that rewrites the feature check function to enable minary view of these retrieved cases. The decision rules(e.g, RCBR to manage the HCA problem. As in traditional CBr, the fea- association rules ) which are generated by data mining techniques ture-check-function is a pre-defined function that returns the simi- further help decision makers to grasp the refined knowledge. Final- larity gap of a particular feature between the target and the case. ly, the successful problem-solving experiences are added to the However, in this case, if PMs query the case base using HCA, then case base for future use. the features are displayed hierarchically, as in Fig. 2. Therefore, the Fig 6 depicts the core algorithm of RCBR. Each feature of a par feature-check-function of traditional CBR should be modified ticular case would be compared the gap with target by relatively accordingly. Restated, as shown in Fig 4, if the feature level is more feature-check-function as Eq (1). The summary of the feature gap than 1(level (fet)>1), then the value of fet-check-function should be in Eq (2)then evaluates the similarity between the target and each replaced by the next level and standardized important weight. ase in the case base. On line(9). the fittest case(s) are retrieved Thus, the fet-check-function would be rewritten by fet-ckf-Rewrite using Reuse mechanism (in Fig. 7)as recommendation cases for di- function. For illustration, for the HCa problem with three level rect reuse by PMs. Additionally, the retrieved case(s) are further requirements displayed in Fig. 2, its similarity function is replaced analyzed by the kdd (knowledge discovery) mechanism of RCBR by level 3 to level 2, and finally to level 1, as shown presented in stage ll, through association mining and statistical analysis On line the dash block area in Fig. 5. The function fet_ckf_Rewrite returns (10), the KDD mechanism generates potential knowledge rules, the revised feature check for the rCBr core algorithm to evaluate and then provides case revision direction. Finally, as revealed in the similarity between the case and the target. Fig. 7, after sorting similarities of cases, the case(s)with the small- Finally, in the stage ll of RCBR, as denoted by a dash-line block est gap are sent to the Pms for further reuse. These retrieved cases area in Fig 3, statistical and mining techniques are adopted to per- are further considered as inputs of RCBR stage ll for KDD analysis. form multiple cases analysis. Since the retrieved cases and target both satisfy the same conditions, the discovered knowledge results 4. Experiment result are useful to decision makers. This study improves the outcomes of traditional CBr by providing descriptive statistical information and An experiment was performed to verify the proposed RCBr. A decision rules. The descriptive statistic information (e.g, the pro- software consultant company in Taiwan (named Company a Function fet_check function= fet_ckf_ Rewritedfet) if( level(fet)>1) fet_check function= wgt e, X fet_ckf_Re write(next_level( fet) else fet_check function= wgt e x fet_ check_function end if; Fig 4. Similarity evaluated for HCA probler
Fig. 4 depicts a recursive algorithm (named fet_ckf_Rewrite mechanism) that rewrites the feature check function to enable RCBR to manage the HCA problem. As in traditional CBR, the feature-check-function is a pre-defined function that returns the similarity gap of a particular feature between the target and the case. However, in this case, if PMs query the case base using HCA, then the features are displayed hierarchically, as in Fig. 2. Therefore, the feature-check-function of traditional CBR should be modified accordingly. Restated, as shown in Fig. 4, if the feature level is more than 1 (level (fet) > 1), then the value of fet-check-function should be replaced by the next level and standardized important weight. Thus, the fet-check-function would be rewritten by fet-ckf-Rewrite function. For illustration, for the HCA problem with three level requirements displayed in Fig. 2, its similarity function is replaced by level 3 to level 2, and finally to level 1, as shown presented in the dash block area in Fig. 5. The function fet_ckf_Rewrite returns the revised feature check for the RCBR core algorithm to evaluate the similarity between the case and the target. Finally, in the stage II of RCBR, as denoted by a dash-line block area in Fig. 3, statistical and mining techniques are adopted to perform multiple cases analysis. Since the retrieved cases and target both satisfy the same conditions, the discovered knowledge results are useful to decision makers. This study improves the outcomes of traditional CBR by providing descriptive statistical information and decision rules. The descriptive statistic information (e.g., the proportion distribution) enables decision makers to capture the preliminary view of these retrieved cases. The decision rules (e.g., association rules), which are generated by data mining techniques, further help decision makers to grasp the refined knowledge. Finally, the successful problem-solving experiences are added to the case base for future use. Fig. 6 depicts the core algorithm of RCBR. Each feature of a particular case would be compared the gap with target by relatively feature-check-function as Eq. (1). The summary of the feature gap in Eq. (2) then evaluates the similarity between the target and each case in the case base. On line (9), the fittest case(s) are retrieved using Reuse mechanism (in Fig. 7) as recommendation cases for direct reuse by PMs. Additionally, the retrieved case(s) are further analyzed by the KDD (knowledge discovery) mechanism of RCBR stage II, through association mining and statistical analysis. On line (10), the KDD mechanism generates potential knowledge rules, and then provides case revision direction. Finally, as revealed in Fig. 7, after sorting similarities of cases, the case(s) with the smallest gap are sent to the PMs for further reuse. These retrieved cases are further considered as inputs of RCBR stage II for KDD analysis. 4. Experiment result An experiment was performed to verify the proposed RCBR. A software consultant company in Taiwan (named Company a • Feature • Weight Case Base Start Input Target Stage I Retrieved case (s) Statistical Information Stage II Retrieved case (s) Knowledge Base Rules Discovered Knowledge Reuse Revised Reuse Retain Reuse Fig. 3. Proposed RCBR model. Fig. 4. Similarity evaluated for HCA problem. H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945 8941
H -L Yang C-S. Wang/Expert Systems with Applications 36(2009)8938-8945 Level 1: Sim(Ci. Target)=WI. CT ckf function+W2. P/T ckf function+ W3. P/T Kw cofunction I: Level 2 to Level Level 2. 1 Level 3 to Leve/2 CT ckf function=Wu. pt ckf finction+WI. cap ckffunci P/T ckF function= W2l. BT ckf function+W22. dur function PT KW check function=W3/. HR ckf function+W32. acc ckf function+ W33. man ckf function+W 34. DSS ckf function Level 3 cap_ck function=W121. HW-ckf function+ Win,. SW_ckf finction ;-- Fig. 5. Illustration for feature-check- function rewrite mechanism. For i=I to n For j=l m if level(fer)>1) fet, _check function=fet_ckf Rewrite( fet,) end if, gap(fer/, fer)=wgt ea, Xfet, -check_function eq(1) nextJ; Sim(Ci, T)=2 gap(fer/', fer) q,(2) Recommendation Co es= line(9) Potential_ Rules, Sol_ Suggestion)=KDD(Recommendation_Cases); line(10) ext l Fig. 6. Core algorithm(I)of RCBR Function Recommendation_Cases=Reuse(Sim, q); New_ sim=sort(sim) q Recommendation Cases=Minimiz e(New sim) Fig. 7. Core algorithm (li)of RCBR. herein )agreed to provide three years of Ir/is project data. The data comprised 54 fields, of which 43 projects remained after perform organization and properties), project dimension(si ing data cleaning. Each project was described in terms of three finished on time and staff involved) and project key
herein) agreed to provide three years of IT/IS project data. The data comprised 54 fields, of which 43 projects remained after performing data cleaning. Each project was described in terms of three dimensions, namely customer dimension (such as customer name, public organization and properties), project dimension (such as budget, finished on time and staff involved) and project keywords Fig. 5. Illustration for feature-check-function rewrite mechanism. Fig. 6. Core algorithm (I) of RCBR. Fig. 7. Core algorithm (II) of RCBR. 8942 H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945
H.-L Yang C-S. Wang/Expert Systems with Applications 36(2009)8938-8945 describing the project content(such as HR project, salary and For instance, consider the situation where a PM tried to arrange insurance policy ) In this experiment, the PMs entered conditions a project plan related to the hr management and staff insurance to query the project case base for planning recommendations. policy transfer (where the importance weight of each project keyword is set to 2)for a bank customer B(the importance weight 4.1. Case retrieved for project planning recommendation: RCBR stage I of customer similarity is set to 3). The PM set the similarity thresh old above 0.6 for references. Sixteen cases were retrieved for As indicated in Fig. 8, PMs entered queries parameters via the references. web interface, and assigned relative importance weights. The Moreover, the customer might want to control the budget be- PMs described their project queries in terms of HCA In Fig. 8a, low NTS200,000(with an importance weight of 5). Hence, the PMs could input their queries via three dimensions: customer, pro- PM could click the"re-submit"button on the retrieved result page. ject keyword and project content. Each dimension can be drilled as indicated in Fig 9, to revise the query. According to the PMs own to present PMs requirements the required level of detail is new query condition, only six cases are retrieved for references. eached as shown in Fig. 8b The PM can click any retrieved case to check these past project Software Proect Construction Recommender System e Me t r nhy eaters ce Company 以 exidia csd如 L i Your Best Parber Al it elizas e pleaded tber Afe od imput you menem Te leoma nmen woul roone yo te LLADAR DAm heerin圆 CBk tda ITing Dotem e时 she duty pourcuroners hmed cxmbrext B曰如四团D 时cpm 出款 Flase cick subnet. Then mmm would aowe you be tn ase for referenc. 加面山购 Fig 8. A Demonstrated experiment interface. Case 3 Project Name: Insurance Policy Tranafer and salary UpgRade Ise mO ans 1. Case 3(Similar 0.83) 2. Case 5(Similarity 0.8) 3. Case 11(Similarity 0.68) 4. Case 8(Similarity 0, 62) 5. Case 12 (Smarty 0.6) You can chick on each case for project construction reference Further, You can press re-gubmet for query revson Or you can cock on refined informaton for further nformaton Re- Subnit Query Fig 9. A Demonstrated experiment result of RCBR stage l
describing the project content (such as HR project, salary and insurance policy). In this experiment, the PMs entered conditions to query the project case base for planning recommendations. 4.1. Case retrieved for project planning recommendation: RCBR stage I As indicated in Fig. 8, PMs entered queries parameters via the web interface, and assigned relative importance weights. The PMs described their project queries in terms of HCA. In Fig. 8a, PMs could input their queries via three dimensions: customer, project keyword and project content. Each dimension can be drilled down to present PMs’ requirements the required level of detail is reached, as shown in Fig. 8b. For instance, consider the situation where a PM tried to arrange a project plan related to the HR management and staff insurance policy transfer (where the importance weight of each project keyword is set to 2) for a bank customer b (the importance weight of customer similarity is set to 3). The PM set the similarity threshold above 0.6 for references. Sixteen cases were retrieved for references. Moreover, the customer might want to control the budget below NT$200,000 (with an importance weight of 5). Hence, the PM could click the ‘‘re-submit” button on the retrieved result page, as indicated in Fig. 9, to revise the query. According to the PMs0 new query condition, only six cases are retrieved for references. The PM can click any retrieved case to check these past project Fig. 8. A Demonstrated experiment interface. Fig. 9. A Demonstrated experiment result of RCBR stage I. H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945 8943
H -L Yang C-S. Wang/ Expert Systems with Applications 36(2009)8938-8945 From 6 retrieved cases Follo wing Rule could be found 2. Average revenue rete is 19%6 3. You should assign a sa or SD at least. [Tell me more.I Some rule you could reference 知””” Fig. 10. The Kdd demonstration of RCBR stage Il. details. The query re-submitting process is similar to the what-if analysis in a on system. 4. 2. Refined knowledge for project planning recommendation: RCBR ystematic methodology as the core of a recommender system. A stage ll new problem description, named hierarchical criteria architecture is first presented to enhance traditional CBR. HCA enables PMs to However, a query with vague criteria might retrieve too many address each dimension carefully in order to present project cases(16 cases in the above original example)to produce useful requirements as deep as possible. The project case base can be pro- information, thus causing information overloading. Therefore, fur- vided online, and allows PMs to update project data immediately ther mining can be performed in RCBR stage ll to generate valuable and globally. Additionally, a revised case-based reasoning algo- nformation rithm, RCBR, is developed to resolve case matching while the target For instance, in the second round of the above example, six and cases are presented in HCA format. RCBR combines the tradi- cases(Case No. 3. 5, 6, 8, 11 and 12)were retrieved to construct tional CBR algorithm with the knowledge discovery techniques. roject plan references If PMs click the"provide refined informa- RCBR has two stages. In stage ll, the knowledge mining from those ion"button, then the statistical mining and analysis are performed retrieved cases of stage I should be useful for project planning. to provide the refined knowledge based on the outcomes of stage 1. making the recommendation result more useful, agile and flexible. As demonstrated in Fig. 10, PMs found that the average cost of such Thus, RCBr enhances the quality of project managers'decisions projects was NTS100, 500, and the average revenue rate was Future work should provide the implemented recommender approximately 19%. Additionally, the duration of project imple- system, in which RCBR algorithm is embedded, for in the mentation was about 26 days, and 85% of projects were completed software industry. A field study also should be performed to eval- on time. Finally, these projects could be run with one system ana- uate the success rate of projects, and the satisfaction of PMs lyst or one designer, plus one programmer. The results of association analysis are presented as follows. A References project in which the customer platform was not transferred was nore likely to be finished on time than one in which it was trans- Aamodt, A,&Plaza, E(1994).Case-based ferred, and the complexity was also lower(confidence 0.83 and Is,and system minimal support 0.7). The above knowledge enabled the project Adomavicius. G. Tuzhilin, A.( 2005) Toward the next generation manager could prepare project plan with more confidence, and systems: A survey of the state-of-the-art and possible extensions. IEEE have additional information to bargain with customers. The result of knowledge discovery could give PMs directions for revising pro- elecheanu, R, Pawar, K., Barson,R, Bredehorst, B,& Weber, F(2003). The oning to decision support in new produc ject plans. Thus, RCBR can provide actionable information for PMs. ufacturing Systems, 14(1), 36-45 This knowledge is particularly useful at the highly uncertain stages Cirovic, G-& cekic, z(2002). Case-based ing model applied as a decision of constructing project plans techniques. ACM Computing Surveys, 32(2), 109-143 5 Conclusions Dweir, F T.& Kablan, M. M (2006). A decision support system for software project Dweir. E. T. Kablan. M. M(2006. USing fuzzy decision or the evaluation Case-based reasoning agorithm can feasibly solve ill-defined oject management internal efficiency. Decision Support Systems, 42(2), and unstructured problems, such as project management In par ticularly, PMs typically need previous project experiences during Decision the early stage of constructing project plans. This is a typical CBR Support Systems, 38(2), 305-317
details. The query re-submitting process is similar to the what-if analysis in a decision system. 4.2. Refined knowledge for project planning recommendation: RCBR stage II However, a query with vague criteria might retrieve too many cases (16 cases in the above original example) to produce useful information, thus causing information overloading. Therefore, further mining can be performed in RCBR stage II to generate valuable information. For instance, in the second round of the above example, six cases (Case No. 3, 5, 6, 8, 11 and 12) were retrieved to construct project plan references. If PMs click the ‘‘provide refined information” button, then the statistical mining and analysis are performed to provide the refined knowledge based on the outcomes of stage I. As demonstrated in Fig. 10, PMs found that the average cost of such projects was NT$100,500, and the average revenue rate was approximately 19%. Additionally, the duration of project implementation was about 26 days, and 85% of projects were completed on time. Finally, these projects could be run with one system analyst or one designer, plus one programmer. The results of association analysis are presented as follows. A project in which the customer platform was not transferred was more likely to be finished on time than one in which it was transferred, and the complexity was also lower (confidence 0.83 and minimal support 0.7). The above knowledge enabled the project manager could prepare project plan with more confidence, and have additional information to bargain with customers. The result of knowledge discovery could give PMs directions for revising project plans. Thus, RCBR can provide actionable information for PMs. This knowledge is particularly useful at the highly uncertain stages of constructing project plans. 5. Conclusions Case-based reasoning algorithm can feasibly solve ill-defined and unstructured problems, such as project management. In particularly, PMs typically need previous project experiences during the early stage of constructing project plans. This is a typical CBR behavior. However, the traditional CBR needs to be revised for applying to real project plan preparation. Thus, to help PMs to build project plans, this study proposes a systematic methodology as the core of a recommender system. A new problem description, named hierarchical criteria architecture, is first presented to enhance traditional CBR. HCA enables PMs to address each dimension carefully in order to present project requirements as deep as possible. The project case base can be provided online, and allows PMs to update project data immediately and globally. Additionally, a revised case-based reasoning algorithm, RCBR, is developed to resolve case matching while the target and cases are presented in HCA format. RCBR combines the traditional CBR algorithm with the knowledge discovery techniques. RCBR has two stages. In stage II, the knowledge mining from those retrieved cases of stage I should be useful for project planning, making the recommendation result more useful, agile and flexible. Thus, RCBR enhances the quality of project managers’ decisions. Future work should provide the implemented recommender system, in which RCBR algorithm is embedded, for usage in the software industry. A field study also should be performed to evaluate the success rate of projects, and the satisfaction of PMs. References Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational issues, methodological variations, and system approaches (Vol. 7: 1). AI Communications, IOS Press. Adomavicius, G., & Tuzhilin, A. (2005). Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE Transaction on Knowledge and Data Engineering, 17(6), 734–749. Belecheanu, R., Pawar, K. S., Barson, R. J., Bredehorst, B., & Weber, F. (2003). The application of case based reasoning to decision support in new product development. Integrated Manufacturing Systems, 14(1), 36–45. Cirovic, G., & Cekic, Z. (2002). Case-based reasoning model applied as a decision support for construction projects. Kybernete, 31(5/6), 896–908. Coello, C. A. (2000). An updated survey of GA-Based multiobjective optimization techniques. ACM Computing Surveys, 32(2), 109–143. Dweiri, F. T., & Kablan, M. M. (2006). A decision support system for software project management. IEEE Software, 67–75. Dweiri, F. T., & Kablan, M. M. (2006). Using fuzzy decision making for the evaluation of the project management internal efficiency. Decision Support Systems, 42(2), 712–726. Garcia, M., Quintales, L., Penalvo, F., & Martin, M. (2004). Building knowledge discovery-driven models for decision support in project management. Decision Support Systems, 38(2), 305–317. Fig. 10. The KDD demonstration of RCBR stage II. 8944 H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945
H.-L Yang C-S. Wang/Expert Systems with Applications 36(2009)8938-8945 R Penalva, F& Bonilla, M. (2008). An association rule mining Rodriguez-Repiso, L, Setchi, R,& Salmeron, J. (2007). Modeling Ir project ng the impact of project management policies on software success with fuzzy cognitive maps. Expert Systems with Applications, 32(2), Han, W. M,& Huang S I(2007) An empirical analysis of risk components and Stamelos. L. Angelis, L(2001). Managing uncertainty in project portfolio cost perfomance on software project. The Journal of System and Sofrware. 8o0, esc b goppenberorat i n& Frolick 04 200 m d eat tesr esetaorsg Tep or. Lee,J,& Lee, N.(2006 Least I cation principle for case-based reasoning: A oftware project planning experience. Expert Systems with Applications, 30(2), Tina, Q, Ma, ],& Liu, Q.(2002). A hy stems with Applications, 23(3). 265-271 Wallace, L, Keil, M.,& Rai, A (2004). Understanding software project risk: A cluster 474)52 project management office. journal of Computer Information reas Yang. H-L&Wang. C S(2008). Two stages of case-based reasoning-integrating rica study Yang, Q- Yin,J- Lin, C,& Pan, R.(2007). Extracting actionable knowledge form on tree. IEEE Transaction on Knowledge and Data Engineering, 19(1) International Joumal of Project Management, 26, 212-220 and project success. of their impact on p
Garcia, M., Roman, I. R., Penalvo, F., & Bonilla, M. (2008). An association rule mining method for estimating the impact of project management policies on software quality, development time and effort. Expert Systems with Applications, 34(1), 522–529. Han, W. M., & Huang, S. J. (2007). An empirical analysis of risk components and performance on software project. The Journal of System and Software, 80(1), 42–50. Lee, J., & Lee, N. (2006). Least modification principle for case-based reasoning: A software project planning experience. Expert Systems with Applications, 30(2), 190–202. Love, E. D., & Irani, Z. (2003). A project management quality cost information system for the construction industry. Information and Management, 40(7), 649–661. Mahaney, R., & Lederer, A. (2003). Information systems project management: An agency theory interpretation. Journal of Systems and Software, 68(1), 1–9. Martin, L., Pearson, J., & Furumo, K. (2007). IS project management: Size, practices and the project management office. Journal of Computer Information System, 47(4), 52–60. Raymond, L., & Bergeron, F. (2008). Project management information systems: An empirical study of their impact on project managers and project success. International Journal of Project Management, 26, 212–220. Rodriguez-Repiso, L., Setchi, R., & Salmeron, J. (2007). Modeling IT project success with fuzzy cognitive maps. Expert Systems with Applications, 32(2), 543–559. Stamelos, I., & Angelis, L. (2001). Managing uncertainty in project portfolio cost estimation. Information and Software Technology, 43(13), 759–768. Standish Group International, INC. (2004). 2004 Third quarter research report. Tesch, D., Kloppenbborg, T., & Frolick, M. (2007). IT project risk factors: The project management professionals perspective. Journal of Computer Information System, 47(4), 61–69. Tina, Q., Ma, J., & Liu, Q. (2002). A hybrid knowledge and model system for R&D project selection. Expert Systems with Applications, 23(3), 265–271. Wallace, L., Keil, M., & Rai, A. (2004). Understanding software project risk: A cluster analysis. Information and Management, 42(1), 115–125. Yang, H.-L., & Wang, C. S. (2008). Two stages of case-based reasoning — integrating genetic algorithm with data mining mechanisms. Expert Systems with Applications, 35(1/2), 262–272. Yang, Q., Yin, J., Lin, C., & Pan, R. (2007). Extracting actionable knowledge form decision tree. IEEE Transaction on Knowledge and Data Engineering, 19(1), 43–56. H.-L. Yang, C.-S. Wang / Expert Systems with Applications 36 (2009) 8938–8945 8945
