Expert Systems with Applications 41 (2014) 3955–3964

Contents lists available at ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier .com/locate /eswa

Optimizing parameters of support vector machine using fast messygenetic algorithm for dispute classification

0957-4174/$ - see front matter � 2013 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.eswa.2013.12.035

⇑ Corresponding author. Address: Department of Civil and ConstructionEngineering, National Taiwan University of Science and Technology, 43, Sec. 4,Keelung Rd., Taipei 106, Taiwan, ROC. Tel.: +886 2 2737 6321; fax: +886 2 27376606.

E-mail address: jschou@mail.ntust.edu.tw (J.-S. Chou).

Jui-Sheng Chou ⇑, Min-Yuan Cheng, Yu-Wei Wu, Anh-Duc PhamDepartment of Civil and Construction Engineering, National Taiwan University of Science and Technology, Taiwan, ROC

a r t i c l e i n f o a b s t r a c t

Keywords:Classification modelHybrid intelligenceOptimizationSupport vector machineFast messy genetic algorithmDispute predictionProject management

Hybrid system is a potential tool to deal with construction engineering and management problems. Thisstudy proposes an optimized hybrid artificial intelligence model to integrate a fast messy genetic algo-rithm (fmGA) with a support vector machine (SVM). The fmGA-based SVM (GASVM) is used for early pre-diction of dispute propensity in the initial phase of public–private partnership projects. Particularly, theSVM mainly provides learning and curve fitting while the fmGA optimizes SVM parameters. Measures interm of accuracy, precision, sensitivity, specificity, and area under the curve and synthesis index are usedfor performance evaluation of proposed hybrid intelligence classification model. Experimental compari-sons indicate that GASVM achieves better cross-fold prediction accuracy compared to other baselinemodels (i.e., CART, CHAID, QUEST, and C5.0) and previous works. The forecasting results provide the pro-active-warning and decision-support information needed to manage potential disputes.

� 2013 Elsevier Ltd. All rights reserved.

1. Introduction

Data mining (DM) has attracted great scientific interest and hasbecome an important research area. Major DM methods includepredictive modeling (supervised learning in machine learning,i.e., classification and regression problems), clustering and associa-tion (unsupervised learning), evolution, pattern matching, datavisualization and meta-rule guided mining (Liao, Chu, & Hsiao,2012). In particular, classification is one of the important missionsin data mining (Ngai, Xiu, & Chau, 2009). The DM- and AI-based ap-proaches are related to computer system programs that attempt toresolve problems by emulating human brain processes (Garg, Rani,& Sharma, 2014; Hajdasz, 2014; Irani & Kamal, 2014; Khashei,Zeinal Hamadani, & Bijari, 2012). Therefore, the use of AI-basedmodels is a potential tool to deal with classification problems inconstruction engineering and management.

Notably, a global trend has become interested for financial pub-lic investment via Public Private Partnership (PPP) which is a finan-cial strategy for stimulating private investments in public works.PPP have proved to be a useful and beneficial instrument. However,PPP project involve a variety of organizations and a large numberof partnerships as well as exist high risk and uncertainty. Disputebetween parties to projects are of great concern to the constructionindustry (Fenn, Lowe, & Speck, 1997; Tang, Shen, & Cheng, 2010).

These problems are also influenced by highly variable and unpre-dictable factors. Because of these difficulties and the importanceof enhancing forecast capability, algorithms with complexityapproaching that of integrated models have been developed to im-prove modeling accuracy, effectiveness and speed (Chou & Lin,2013; Donis-Díaz, Muro, Bello-Pérez, & Morales, 2014; Pai, Hung,& Lin, 2014; Seera & Lim, 2014; _Ilhan & Tezel, 2013). However,enhancing the generational capability of advanced DM techniquesis still in need for the PPP-related projects.

Taiwan has legally supported PPP projects for more than tenyears. The Taiwan Public Construction Commission (TPCC) has ac-tively promoted and encouraged private-sector participation ininfrastructure and building construction throughout Taiwan.According to the TPCC, the dispute rate was 23.6% during 2002–2009. The most common processes for handling disputes are medi-ation or non-mediation (e.g., arbitration, litigation, negotiation, andadministrative appeals) procedures. In Taiwan, up to 84% of PPPprojects are settled by mediation or negotiation within only1–9 months after disputes occurs (PCC., 2010). Notably, arbitrationor litigation costs all parties considerably more time and moneywhen a mediated agreement cannot be reached. To address thesechallenges, the proposed classification approaches predict propen-sity for project claims providing supportive information needed bygovernmental authority to furnish contact documents in the pre-paratory and bidding phases of PPPs.

Moreover, the dispute between PPP participantes commonlyoccur unexpectedly and may involve many issues, including suretybond issue, sub-contractor qualifications, licenses, permits, invest-ment scale, resident rights, government guarantees, excessive

3956 J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964

profits, operating period, taxation, and default loan commitment(Jones, 2006). Numerous studies show that an efficient, effective,and fair dispute resolution process is essential for PPP project suc-cess (Cheung, 1999; Cheung, Suen, & Lam, 2002; Cheung, Yiu, &Chan, 2010; Chou & Lin, 2013; Goetz & Gibson, 2009; Menassa &Peña Mora, 2010). Therefore, development of intelligence modelscan enable early warming of potential dispute resolutions priorto project initiation to be becoming crucial.

Several studies have attempted to develop the hybrid AI models(Hsiao, Wang, Wang, Wen, & Yu, 2012; Irani & Kamal, 2014;Khashei et al., 2012; Lee, Rajkumar, Lo, Wan, & Isa, 2013; Seera &Lim, 2014) by combining one with other techniques to enhancetheir performance results, such as combination of genetic algo-rithm and support vector machine (Fei & Zhang, 2009; Huang &Wang, 2006; Jiao & Liu, 2011; Zhao, Fu, Ji, Tang, & Zhou, 2011; _Ilhan& Tezel, 2013). Particularly, the mechanism for setting tuningparameters in AI models is an important issue that is widely recog-nized by scholars in many different disciplines (Donis-Díaz et al.,2014; J. Huang, Bo, & Wang, 2011; Pai et al., 2014; _Ilhan & Tezel,2013). In practice, identifying the best set of parameters for a mod-el is an optimization problem. Therefore, integrating AI-basedmodels with nature-inspired algorithms has been proposed as asolution to the above problems.

Advanced AI-based approaches include support vector machine(SVM), which is a machine learning technique with advanced fea-tures that enable good generalization and fast computation. TheSVM has been demonstrated very powerful in solving classificationproblems (Lin, 2002). Although the SVM model has proven highlyeffective for solving classification problems, it has major draw-backs (Jiao & Liu, 2011). The accuracy of the SVM model dependson parameters set in advance. For such need, fast messy geneticalgorithms (fmGA)-based method developed by Goldberg, Deb,Kargupta, and Harik (1993) is known for its flexibility in allowinghybridization with other methodologies to obtain enhanced solu-tions (David E. Goldberg et al., 1993). The primary difference be-tween an fmGA and other genetic algorithms is its ability tomanipulate explicit building blocks of genetic material to obtainglobal optimization (Hettiarachchi, Noman, & Iba, 2013; Wu,2005). The fmGA can efficiently find the optimal solution of thelarge-scale permutation problems (Knjazew, 2002). These advanta-ges make fmGA logical candidates for overcoming the disadvan-tages of SVM.

The aim of this study, thus, is to employ an auxiliary hybridtechnique which SVM will call fmGA as subroutine to optimizeits structure parameters. The fmGA-based support vector machine(GASVM) enhance their efficacy to early predict PPP disputelikelihood and potential resolutions, thereby alleviating the futureadverse effects of disputes on project delivery, operation, andtransfer from a governmental prospective. To demonstrate the effi-cacy of proposed hybrid system, this study used PPP project datacollected by TPCC and compared well-known classification andregression-based models (e.g., CART, CHAID, QUEST, and C5.0)using classification performance measure in term of accuracy, pre-cision, sensitivity, specificity, area under the curve and synthesisindex. For avoid bias from data (Kohavi, 1995), cross-fold valida-tion was also executed.

The rest of this study is organized as follows. Section 2 thor-oughly reviews AI literature and the application of AI to predictconstruction claims and litigation outcomes. Section 3 character-izes the research methodology, providing a theoretical basic forclassification performance models and elaborates on the GASVMmodel proposed in this study. Section 4 describes and analyzesnumerical example and analytical outcomes using classificationperformance measures. Conclusions are given in Section 5, alongwith directions for further research.

2. Literature review

The hybrid AI and forecasting techniques are widely used invarious engineering and management fields (Hsiao et al., 2012;Huang & Wang, 2006; Kim & Kang, 2012; Kim & Shin, 2007;Moosmayer, Chong, Liu, & Schuppar, 2013), and have been demon-strated to enhance overall performance. Nevertheless, their effec-tiveness and efficiency are rarely applied in the constructionindustry, particularly in public–private partnership (PPP) project do-main. To support the dispute resolution process, several varieties oftools and systems have been developed (Chong, Mohamad Zin, &Chong, 2012; El-Adaway & Kandil, 2010; Ilter, 2012; Jin & Zhang,2011; Kassab, Hegazy, & Hipel, 2010; Seifert, 2005). Applying thesetechniques is useful for both researchers and practitioners to betterunderstand the complex nature of PPP project. Since a disputealways takes in numerous complex and interconnected factors thatare difficult to rationalize, using DM techniques is now among themost effective methods for improving prediction accuracy relatedto some cases such as construction litigation (Chau, 2006; Chau,2007; Pulket & Arditi, 2009a; Pulket & Arditi, 2009b); constructionprocurement litigation (Arditi & Pulket, 2010); and change-order-triggered disputes (Chen, 2008; Chen & Hsu, 2007).

Arditi and Tokdemir (1999) have developed several models onthe same dataset using AI techniques to enhance prediction resultin conventional construction procurement litigation as a predictionaccuracy of 83.33% was achieved with a case-based reasoning(Arditi & Tokdemir, 1999), 89.95% was achieved with boosted deci-sion trees (Arditi & Pulket, 2005), and 91.15% was attained withintegrated prediction modeling (Arditi & Pulket, 2010). Moreover,their studies have attempted to minimize the number of construc-tion litigation cases by using neural network to predict the likelycourt rulings and obtained a prediction rate of 67% for litigationoutcomes (Arditi, Oksay, & Tokdemir, 1998). They argued that ifoutcome of construction litigation can be predicted with higheraccuracy and reliability by using these approaches, all parties in-volved in the construction process could save considerable money,time, and aggravation.

Furthermore, Chau (2006) found that excluding the above casestudies, AI techniques are rarely applied in the legal field (Chau,2006). Thus, Chau utilized AI techniques based on particle swarmoptimization (PSO) to predict construction litigation outcome, afiled in which relatively new DM techniques are rarely applied.The PSO-based ANN technique developed by Chau obtained therate of prediction is up to 80%, which is much higher than chance.However, Chau suggested using additional case factors related tocultural, psychological, social, environmental, and political featuresin the future work.

The other AI techniques were used for construction dispute. Incase of changing orders of construction process and design, Chen(2008) developed a K nearest neighbor based knowledge sharingmodel, which obtained 84.38% accuracy in predicting lawsuitsbased on a nationwide study of US court records (Chen, 2008).Chen and Hsu (2007) further employed hybrid ANN-CBR modelwith disputed change order dataset to achieve early-warm infor-mation. The ANN classifier demonstrated comparable predictionaccuracy (84.61%) (Chen & Hsu, 2007). In case of dispute settle-ment, Cheng, Tsai, and Chiu (2009) refined and improved the con-ventional CBR approach by combining fuzzy set theory with a newsimilarity measurement that integrates Euclidean distance and co-sine angle distance (Cheng et al., 2009). Their model successfullyextracted the knowledge and experience of experts from 153historical construction dispute cases collected manually frommultiple sources.

Several studies have demonstrated that hybrid AI schemesgenerate promising results in many industries (Chen & Hsu,

J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964 3957

2007; Donis-Díaz et al., 2014; Hettiarachchi et al., 2013; Kim &Shin, 2007; Lee, 2009; _Ilhan & Tezel, 2013). For instance, Kim, Yoon,An, Cho, and Kang (2004) applied the back-propagation network(BPN) model incorporating GA in optimizing both the neural net-work size and its parameter (Kim et al., 2004). The combined mod-el showed that was more effective and accurate in estimatingconstruction costs than BPN model using trial and error. Wu,Tzeng, and Lin (2009) developed a hybrid GASVM model to forecastthe maximum electrical daily load by using GA to optimize kernelfunction and kernel parameters of SVM. This model outperformedany other model employed in the European Network on IntelligentTechnologies for Smart Adaptive Systems network (Wu et al.,2009).

Similarly, Chen (2007) employed GASVM model to predict en-gine system reliability. To build an SVM model efficiently, modelparameters were optimized as regularization parameter C, band-width of the kernel function r2, and the tube size of e-insensitiveloss function. The experimental results demonstrated that GASVMmodel forecast more precise than ANN model and traditional auto-regressive integrated moving average model (Chen, 2007). More-over, the model proposed by Hsiao et al. (2012) integrated thecomponent ratios method, fuzzy adaptive learning control network,fmGA, and three-point cost estimation method to solve cost-esti-mating problems under conditions of limited and uncertain data(Hsiao et al., 2012). In this sense, hybrid approaches are considereda promising research area in the near future (Seera & Lim, 2014).

Additionally, Chau (2006) utilized a split-step PSO algorithmwhich is employed to train multi-layer perceptions for predictionof the outcome of construction litigation (Chau, 2006), its perfor-mance is much better than the benchmark backward propagationalgorithm and conventional PSO algorithm. Using the same datasetin this study, Chou and Lin (2013) proposed an ensemble approachby combining best models to predict dispute propensity in PPPprojects (Chou & Lin, 2013). Their study demonstrated that the pro-posed ensemble model (i.e., SVM + ANNs + C5.0) was more accu-rate than single models with prediction accuracy of 84.33%.

Generally, most of above related works focused on either spe-cific change-order disputes or conventional contracting projects.Management personnel typically is beneficial when the taskforcehas a decision-support tool for forecasting dispute propensityand for early planning of how disputes should be resolved beforeproject initiation (Marzouk, El-Mesteckawi, & El-Said, 2011). How-ever, prediction problems are often complex because they involvesubstantial uncertainty, vagueness, and incomplete or inexact data.Therefore, the inference process must fit environmental conditions(Mareels & Polderman, 1996).

Since humans can process and solve complex problems, eventhose involving uncertainty, imprecision, and incomplete informa-tion, imitating human inference is an effective approach forpredicting disputes. Characteristics and environments for con-struction projects under the PPP strategy differ significantly fromthe contractor-owner relationships which require effective toolsvia hybrid AI algorithms. Thus, the proposed model in this studycombines fmGA and SVM (GASVM), which is still open to tech-niques in the PPP domain knowledge, to possibly provide improveddispute classification accuracy. Modeling performance is thencompared with several well-known classification models and pre-vious works.

3. Research methodology

3.1. Classification-based models

When response variable is categorical rather than continuous,prediction problems become data classification problems.

Classification techniques are based on learning from examples thatmap input vectors into one of several desired output classes. Thiswork used four classification-based techniques as baseline models(i.e., CART, Exhaustive CHAID, QUEST, and C5.0) for automaticallycreating and comparing default models of binary numerical out-comes. Moreover, these techniques use different learning mecha-nisms that are worth of investigating in the case of disputeclassification. Default values were set in numerical predictor nodesusing the IBM SPSS modeler (IBM., 2010), a powerful and versatiledata analytics workbench, to develop the classification and regres-sion-based models in predicting PPP project dispute.

3.1.1. CART and QUESTClassification and regression tree (CART) technique recursively

partitions data into increasingly homogenous subsets (Breiman,Friedman, Olshen, & Stone, 1984). Each of the two subsets is recur-sively split until the homogeneity criterion or some other stoppingcriterion is met. Decision tree methods are far superior to othermodeling techniques in terms of logic rules. A relatively new bin-ary-split decision tree algorithm used for data classification isQUEST (Quick, Unbiased and Efficient Statistical Tree). The QUESTalgorithm resembles CART except that QUEST uses an unbiasedvariable selection technique by default and uses imputation in-stead of surrogate splits to compensate for missing values. There-fore, QUEST is suitable for selecting categorical predictorvariables with multiple categories (Loh & Shih, 1997).

3.1.1. C5.0Another classification technique recently developed by Quinlan

(2007) is C5.0 (Quinlan, 2007). The decision trees were obtained bythe greedy algorithm use boosting technology to improve accuracyin identifying samples. The top-down approach (divide-and-con-quer) to decision tree induction starts with a training set of tuplesand their associated class labels. The tree is constructed by recur-sively partitioning the training set into smaller subsets (Tan, Steinbach,& Kumar, 2006). The main difference between CART and C5.0 isthat the former performs only binary splits, which gives binarytrees, whereas the latter performs a split for each category, whichgives a ‘‘bushlike’’ structure. A good alternative to limiting treegrowth is pruning the full-grown tree. The CART and CART-likeprocedures use validation data to prune deliberately overgrowntrees by using training data whereas C5.0 uses training data forboth growing and pruning the tree (Shmueli, Patel, & Bruce, 2007).

3.1.2. Exhaustive CHAIDExhaustive Chi-squared Automatic Interaction Detector

(Exhaustive CHAID) avoids over-fitting the full-grown tree to thetraining data by continuously merging predictor categories untilonly two super categories remain. The algorithm also uses a recur-sive partitioning method that predates CART technique and iswidely applied in diverse domains (Shmueli et al., 2007). It testsfor independence by using Chi-square test to assess whether split-ting a node obtains significantly improved purity. Particularly, thepredictor with the strongest association (according to p-value)with the response variable at each node is used as a split node. Ifthe tested predictor does not show a statistically significantimprovement, no split is performed, and the algorithm stops.

This study, however, proposes the use of Exhaustive CHAID,which was developed to address the limitations of the CHAID tech-nique (Biggs, Ville, & Suen, 1991), to classify the target field. Specif-ically, CHAID may sometimes fail to optimize the split for apredictor variable since it stops merging categories as soon as itfinds that all remaining categories significantly differ. ExhaustiveCHAID avoids this by continuously merging predictor categoriesuntil only two super categories remain. It then identifies the pre-dictor in the series of merges and the set of categories that gives

3958 J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964

the strongest association with the target variable and computes anadjusted p-value for that association. Thus, Exhaustive CHAID findsthe best split for each predictor and chooses which predictor tosplit by comparing their adjusted p-values (SPSS., 2007).

3.2. GA-based SVM model

This section presents ideas regarding the GASVM in terms of itsmethodology and flowchart. Fig. 1 shows how the synergisticstructure joining the fmGA to the SVM-based classification enablesthe SVM to identify complex mapping relationships between in-puts and outputs. It classifies data using different class labels bygenerating a support vector set, which contains members of thetraining input set that outlines a hyperplane in a feature space.

It also provides a generic mechanism that uses a kernel functionto fit the hyperplane surface to the training data. The user may se-lect a kernel function (e.g., linear, polynomial, radial basis, or sig-moid) for the SVM during the training process, which identifiessupport vectors along the function surface. Previous study sug-gested that radial basis function (RBF) is generally a reasonablefirst choice (Hsu, Chang, & Lin, 2003). Unlike the linear kernel,the RBF maps samples nonlinearly into a higher dimensional spacethat usually yields more promising results compared to other ker-nels (Hsu & Lin, 2002; Lin & Lin, 2003). Therefore, the RBF is ap-plied to construct SVM as the kernel function.

Since the SVM requires users to set optimal parameters, SVMparameters must be obtained simultaneously. For optimal SVMprediction accuracy, parameters that should be optimized includepenalty parameter C and kernel function parameters such as thec of the RBF kernel. One proposed alternative to finding the bestC and c when using the RBF kernel function is grid algorithm. How-ever, this method is time consuming and performs poorly. Thus theobjective of this proposed hybrid model is to use the fittest shapesof SVM with minimum number of support vectors and optimalSVM parameters to preserve acceptable classification. The opti-mized classification algorithm is briefly introduced below.

3.2.1. Fast messy generic algorithmThe fmGA is characterized by its relative immunity to high

dimensionality and local minima and its potential use in hybridSVMs. These advantages make fmGA a logical candidate foraddressing the disadvantages of SVM. Unlike the well-known sim-ple genetic algorithm, which uses fixed length strings to representpossible solutions, fmGA forms strings of varying length frommessy chromosomes. The fmGA can also optimize solutions effi-ciently in large-scale permutation problems. For example, it cansimultaneously optimize SVM parameters C and c.

FmGA and SVM have proven effective in solving various projectmanagement problems. Considering the characteristics and merits

Optimal C, γ

Terminationon check

Project dispute

prediction model

GAOptimization

Yes

No

Fig. 1. GA-based S

of each, the two were combined in the proposed model, i.e., GAS-VM. In the GASVM, the SVM primarily addresses learning andcurve fitting while fmGA addresses parameter optimization. Thismodel was developed to obtain the C and c parameters with min-imal classification error.

3.2.2. Support vector machine-based classificationIn classification problems, SVM identifies a separate hyperplane

that maximizes the margin between two classes. Maximizing themargin is a quadratic programming problem, which can be solvedfrom its dual problem by introducing Lagrangian multipliers. How-ever, searching for a suitable hyperplane in input space is oftenoverly restrictive for practical use. One solution is mapping the in-put space into a higher dimension feature space and then searchingfor the optimal hyperplane. Even without knowledge of mapping,the SVM can still find the optimal hyperplane by using dot productfunctions in feature space, called kernels. The optimal hyperplanecan be expressed as a combination of several input points, calledsupport vectors.

The main purpose of SVM is estimating a classification functionby using input–output training data from two classes (x1, y1), ...,(xn, yn) 2 Rm � {±1}. The goal of classification functions is to establisha hyperplane equation that divides training data and leaves allpoints of the same class on the same side while maximizing the min-imum distances between the hyperplane and each of the two classes(w, b). Where w represents the weight vector realizing a functionalmargin of 1 on the positive point x+ as well as the negative pointx�, and the geometric margin can be computed as follows:

yiðw � xi þ bÞP 1; i ¼ 1; . . . ;m ð1Þ

The optimal hyperplane w � xþ b ¼ 0 is geometrically equiva-lent to maximizing the margin, i.e., the distance between the twoparallel planes w � xþ b ¼ 1 and w � xþ b ¼ �1. The Euclidean

length of the margin is 2/kwk2, where kwk2 ¼Pm

i¼1w2i . The maxi-

mum margin is also the minimum 2-norm kwk2, subject to con-straint (2). Therefore, the problem can be formulated as

minw;b

kwk2

2

subject to yiðw � xi þ bÞP 1ð2Þ

Since classes can rarely be separated linearly, generalizing theoptimal plane problem is needed. Thus, a set of variables n thatmeasures constraint variation is added for each point. The final for-mulation is

minw;b;n

kwk2

2 þ Cm

Xm

i¼1

ni

subject to yiðw � xi þ bÞ þ ni P 1ni P 0 i ¼ 1; . . . ;m

ð3Þ

Train SVM model Input data

Classification error

evaluation

VM flowchart.

Start

Initialize competitive template

Outer loopterminate

Inner loopterminate

Probabilisticallyinitialization

Evaluate every individual

Threshold selection

Building blocks filter

Evaluate every individual

Cut and splice

Mutation

Evaluate every individual

Competitive template

Epoch=1

Epoch=Epoch+1 Era=1

Era=Era+1

Yes

Yes

No

No

Primordialphase

Juxtapositionalphase

Initial phase

Primordial phaseterminate

Juxtapositional phaseterminate

No

No

End

Optimal C,

Yes

Yes

Fig. 2. Adaptation process.

J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964 3959

where C is a penalty parameter (error penalty) chosen by the fmGA.Eq. (3) can be solved by the classical Lagrange multipliers and

Karush–Kuhn–Tucker conditions (Vapnik, 1995). The decisionfunction can be written as

f ðxÞ ¼ signðw � xþ bÞ ¼ signXm

i¼1

yiaikðx; xiÞ þ b

!ð4Þ

where ai,b is calculated using training data.Some kernel functions k(xi,xj) include polynomial, radial basis

function (RBF) and sigmoid kernels (Witten & Frank, 2005) (Eqs.(5)–(7), respectively). Kernel parameters in the kernel functionsshould be optimized for maximum predictive accuracy.

Polynomial kernel:

kðxi; xjÞ ¼ ð1þ xi � xjÞd ð5Þ

Radial basis function kernel:

kðxi; xjÞ ¼ expð�ckxi � xjk2Þ ð6Þ

Sigmoid kernel:

kðxi; xjÞ ¼ tanhðkxi � xj � dÞ ð7Þ

Notably, any function k(xi,xj) satisfying Mercer’s condition canbe used as the kernel function. Among these functions, the Gauss-ian function can map the sample set from the input space into ahigh-dimensional feature space effectively, which is good for rep-resenting the complex nonlinear relationship between the inputand output samples (Hsu et al., 2003). Furthermore, the linear ker-nel is a special case of RBF (Keerthi & Lin, 2003) that the linear ker-nel with a penalty parameter C performs the same RBF kernel withthe parameters (C; c).

The sigmoid kernel also behaves similarly to the RBF for certainparameters. However, the sigmoid kernel is not better than the RBFkernel in general (Lin & Lin, 2003). The second reason is that thenumber of hyper parameters affects model selection complexity.The polynomial kernel has more hyper parameters compared tothe RBF kernel. The RBF kernel presents fewer numerical difficul-ties. Moreover, the sigmoid kernel is invalid (i.e., not the innerproduct of two vectors) under certain parameters (Vapnik, 1995).Recent works also suggested the use of RBF kernel function inthe SVM model is appropriate and sufficient (Huang & Wang,2006; Huang et al., 2011; Pai et al., 2014; _Ilhan & Tezel, 2013). Inthis regard, the RBF kernel function is used for nonlinear relation-ships between class labels and attributes in this study.

3.2.3. Integration of fmGA and SVMFig. 2 illustrates the adaptation processes (initial/primordial/

justapositional phases) of integrating fmGA and SVM.

3.2.3.1. Initial phase. Probabilistic initialization. The goal of the ini-tialization phase is to create a population of strings containing allpossible Building Blocks (BBs) of order k. The fmGA performs theso-called ‘‘probabilistically complete initialization’’ process, whichrandomly generates n chromosomes of length cc, where k < cc 6 land l represent problem length. The cc value may be chosenarbitrarily, and is usually defined as l � k. The population size nmay be approximated using Eq. (8) for the binary-coded problem(Goldberg, Deb, & Horn, 1991).

n ¼

l

k

� �l� k

cc � k

� �2cðaÞb2ðm� 1Þ2k ð8Þ

where c(a) represents the square of a normal random deviate corre-sponding to tail-probability a and b represents the signal-to-noise

ratio (i.e., the ratio of fitness deviation to the difference betweentwo competing BBs). Variable m represents BB number and k repre-sents BB order. Those parameters may be set arbitrarily in accor-dance with the problem.

Evaluate individuals. The purpose of evaluation is to assess thefitness of chromosomes. The process describing the GASVM calcu-lates the fitness of each chromosome. In the initial generation, theGASVM evaluates individuals. The aim of the model is to obtain asolution that provides both a high degree of accuracy and the abil-ity to be generalized to a broader problem set. While model accu-racy in terms of input patterns can be improved by increasing thenumber of support vectors, an accurate model that is made to fitinput patterns does not necessarily capture overall problem behav-ior well. In general, such models suffer from input pattern dataover-fitting and a deterioration of generalization properties. Thus,the objective of the fusion model is to use the fittest shapes ofSVM with a minimum number of support vectors and optimal

3960 J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964

SVM parameters to preserve an acceptable level of prediction accu-racy in posed optimization problems.

3.2.3.2. Primordial phase. Threshold selection. The primordial phasefilters out ‘‘bad’’ genes that do not belong to BBs, so that the resul-tant population encloses a high proportion of ‘‘good’’ genes belong-ing to BBs. Two operations, building-block filtering and thresholdselection, are performed during the primordial phase. A thresholdmechanism has been used (Biggs et al., 1991) to restrict competi-tion between building blocks that share little in common, wheretournament selection between two strings is only permitted if theyshare a greater than expected number of genes in common. In ran-dom strings of two different lengths, cc1, cc2, the threshold can becalculated using Eq. (9).

h ¼ ½cc1cc2=l� ð9Þ

Building Blocks Filter. In the initialization procedure described inthe previous section, the string starts off at a length approximatingthe problem length. In order to get the fmGA to function properly,this initial length must be reduced until it measures somethingclose to building-block length k (Biggs et al., 1991). The key to get-ting this to work properly is pumping up sufficient good buildingblock copies so that, even following random deletion, one or morecopies remain for subsequent processing.

3.2.3.3. Juxtapositional phase. Cut and Splice. Messy operators,which include cut–splice and mutation operators, are used as ge-netic operators in the fmGA. The cut–splice operator, similar tothe crossover operator in the simple GA, is used to recombine dif-ferent strings to create new strings. The cut operator breaks amessy string into two parts using a cut probability Pc = Pk(c � 1),where Pk is a specified bit-wise cut probability and c representsstring length. String length correlates positively with the probabil-ity that the string will be cut. The cut point is chosen randomlyalong the string. The splice operator joins two strings with a spec-ified splice probability (Ps). For example, as shown in Fig. 3, twostrings ((2 1)(1 0)(5 0)(3 1)) and ((1 1)(2 0)(6 0)(4 1)) are recom-bined into ((2 1)(1 0)(4 1)) and ((1 1)(2 0)(5 0)(3 1)) after beingcut and spliced. Goldberg et al. (1993) proposed Pk ¼ 2=l and amaximum string length, after being cut and spliced, of 2l, where lrepresents problem length. Ps is typically set to 1 (Goldberg et al.,1993).

Mutation. Mutation produces spontaneous random changes invarious chromosomes, which protects against premature loss ofimportant notations. In the GASVM, mutation serves to adjustthe value of SVM parameters and activation slopes for betterperformance. It alters one or more genes with a probability (pm).The mutation operator perturbs the allele values of the messy

(2 1)(1 0)(5 0)(3 1)

(1 1)(2 0)(6 0)(4 1)

(2 1)(1 0)(4 1)

(1 1)(2 0)(6 0)(5 0)(3 1)

Original

After cut and splice

cut

Fig. 3. Cut–splice operator.

(2 1)(1 0)(6 0)(5 0)(3 1)

(2 1)(1 1)(6 0)(5 0)(3 0)

Original

After mutation

Fig. 4. Mutation.

chromosome by switching them from 1 to 0 and vice versa, witha predefined probability pm (see Fig. 4).

4. Numerical example

4.1. Dispute data collection

To demonstrate the applicability and efficiency of the disputeclassification schemes, this study used PPP project data collectedby the Taiwan Public Construction Commission (TPCC) – the gov-ernmental authority overseeing public services and infrastructureconstruction in Taiwan. The study database contains 584 PPP pro-jects overseen by the TPCC during 2002–2009. Of the 584 surveysissued, 569 were returned completed, for a response rate of 97.4%.This high return rate from various governmental units may beattributable to the TPCC taskforce, the most senior governmentalauthority, conducting survey activity. The questionnaire includeditems to collect social demographic data, background information,project characteristics, and project dispute resolution data.

Several projects had more than one dispute (the highest ratewas nine disputes for one project) during various project phases.Thus, the overall dataset included data for N = 645 cases (i.e.,N2 = 493 non-dispute cases; N1 = 152 dispute cases) when count-ing one dispute occurrence as a single case. Based on expert feed-back and data availability, project attributes and their derivativesthat were clearly relevant to the prediction output of interest wereidentified by survey items.

Table 2 summarizes the statistical profiles of categorical labelsand numerical ranges for the resulting study samples. Out of allPPPs analyzed, 59.5% were performed by the central government.Over the past eight years, most public construction has involvedcultural and education facilities (25.3%), sanitation and medicalfacilities (20.8%), transportation facilities (18.1%), and major tour-site facilities (10.5%). In accordance with the economic planningand development policy, 48.5% of projects were located in northernTaiwan. Industrial departments (38.6%) and service departments(50.7%), which were classified according to standard industry def-initions, comprised most private sector investment. In most cases(91.0%), the government provided land and facility designs to at-tract the investors.

Historically, the three major PPP strategies for delivering publicservices were BOT (23.7%), OT (52.7%), and ROT (23.6%). The WorldBank Group (WBG) (WBG., 2011) defines BOT (Build, operate, andtransfer) as a strategy in which a private sponsor builds and oper-ates a new facility before transferring it to the government at theend of the contract period. The government usually provides reve-nue guarantees through long-term take-or-pay contracts. AnotherPPP strategy defined by the WBG classifications is rehabilitate,operate, and transfer (ROT), in which a private sponsor rehabili-tates an existing facility and then operates and maintains the facil-ity at its own risk for the contract period. Projects involving onlymanagement and lease contracts are classified as OT (operateand transfer) projects.

Further, flagship infrastructure projects refer to those that areimportant and fairly large in scale (i.e., the flagship projects in thisstudy had an average value of approximately 841 million NTD).The collected data indicated that the total amount procured viaPPP approximated 543 billion NTD. The mean capital investmentby the government and private sectors per project was 63.5 millionNTD and 777.8 million NTD, respectively. Notably, average privatecapital investment ratio was as high as 91.4%. The mean durationof licensed facility operations was about 12.0 years (maximum,60 years).

To measure the dependencies between the categorized data,contingency table analyses were compared between the distinct

Table 1Confusion matrix.

Predicted

Positive Negative

Actual Positive tp fnNegative fp tn

J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964 3961

predictors and response variable via Chi-square testing to infer therelationships (Table 3). All tests obtained statistically significantresults with alpha levels of at least 5% except the variable (i.e.,planning and design unit), which was a rejection of the nullhypothesis, i.e., no relationship was observed between the row var-iable (input variable) and the column variable (output variable).

For example, among the disputed cases (N1 = 152), central gov-ernment agencies had a higher probability of encountering dis-putes (67.1% probability) compared to municipal (15.1%) andlocal agencies (17.8%). Particularly, in public construction and facil-ity type Nos. 1, 6, 7, 10, 11, and 20 (Table 2), disputes occurred in76.4% of projects. The data showed that 85.5% disputes occurred innorthern and southern Taiwan.

Interestingly, 92.1% of the disputes occurred when the govern-ment provided the land and was responsible for facility designwhile merely 2% occurred when private investors provided theland and designed the facilities themselves. Among the PPP strate-gies, the probability of disputes was higher in BOT (49.3%) than inOT (32.2%) and ROT (18.4%). Notably, once the project was legallypromoted as major infrastructure, the likelihood of a dispute

Table 2Project attributes and their descriptive statistics.

Attribute Data range, categorical label or statistica

Input variablesType of government agency in charge Central authority (59.5%); municipality (Type of public construction and facility 1: Transportation facilities (18.1%);

2: Common conduit (0%);3: Environmental pollution prevent4: Sewerage (1.1%);5: Water supply facilities (0.5%);6: Water conservancy facilities (2.57: Sanitation and medical facilities8: Social welfare facilities (3.9%);9: Labor welfare facilities (1.2%);

10: Cultural and education facilities11: Major tour-site facilities (10.5%);12: Power facilities (0%);13: Public gas and fuel supply facilit14: Sports facilities (3.3%);15: Parks facilities (2.5%);16: Major industrial facilities (0.5%);17: Major commercial facilities (1.9%18: Major hi-tech facilities (0.2%);19: New urban development (0%);20: Agricultural facilities (5.6%)

Project location North (48.5%); center (21.2%); South (24Executive authority Central authority (36.0%); municipality (Type of invested private sector Standard industry classification-primaryPlanning and design unit Government provides land and plans fac

(5.9%); private provides land and designPPP contracting strategy BOT (23.7%); OT (52.7%); ROT (23.6%)Major public infrastructure/facility Promoted as major public infrastructureProject scale Range: 0–60,000,000; Sum: 5.43E8; mea

about 1:30 as of Apr. 2011)Government capital investment Range: 0–9,600,000; sum: 40,975,392.41Private capital investment amount Range: 0–60,000,000; sum: 5.02E8; meaPrivate capital investment ratio (PCIR) Range: 0–100; mean: 91.4729; standardLicensed operations duration Range: 0–60; mean: 11.9778; standard d

Output variablesDispute propensity No dispute occurred (76.4%); dispute occ

involving PPP (38.8%) was lower than that in non-major infrastruc-ture (61.2%).

Moreover, once the project value exceeded 50 million NTD, thedispute propensity was 4.33 times higher than that for projectsvalued between 5 and 50 million NTD or less than 5 million NTD.However, when the private sector investment exceeded 75%, thedispute tendency increased to 92.8%. Notably, dispute patternswere significantly related to the licensed operation period. Table 3summarizes the statistical results of the cross-analysis.

4.2. Performance measure for classification

Researchers often use k-fold cross-validation algorithm to min-imize bias associated with the random sampling of the training andholdout data samples. Kohavi (1995) further confirmed that ten-fold validation testing obtains the optimal computation time andvariance (Kohavi, 1995). Thus, a stratified ten-fold cross-validationapproach was used to assess model performance in this study.

Classification performance can be evaluated by computing thenumber of correctly recognized class examples (true positives;tp), the number of correctly recognized examples that do not be-long to the class (true negatives; tn), and the number of examplesthat were either incorrectly assigned to the class (false positives;fp) or that were unrecognized as class examples (false negatives;fn) (Sokolova & Lapalme, 2009). The four counts constitute a confu-sion matrix (Table 1), which can generate measures commonlyused for binary classification such as accuracy, precision, sensitiv-ity, specificity, and area under the ROC curve (AUC) (Ferri,Hernández-Orallo, & Modroiu, 2009; Kim, 2010).

l description

11.5%); local government (29%)

ion facilities (2.3%);

%);(20.8%);

(25.3%);

ies (0%);

);

.5%); East (5.3%); isolated island (0.5%)36.1%); local government (27.9%)(0.2%); secondary (38.6%); tertiary (50.7%); quaternary (10.5%)ility (91.0%); government provides land and private investor designs facilitys facility (3.1%)

/facility in PPP act (80.1%); not major infrastructure/facility (19.9%)n: 841337.1776; standard deviation: 3.52061E6 (thousand NTD; USD:NTD is

; mean: 63527.7402; standard deviation: 5.11192E5 (thousand NTD)n: 777809.4374; standard deviation: 3.32433E6 (thousand NTD)deviation: 25.42269 (%)eviation: 13.39007 (year)

urred (23.6%)

Table 3Contingency table and chi-square test results for disputed cases.

Project attributes p-value

Disputeoccurred (%)

Agency .002Central authority 67.1Municipality 15.1Local government 17.8Type of public construction .000Transportation facilities 10.5Water conservancy facilities 9.9Sanitation and medical facilities 17.1Cultural and education facilities 13.2Major tour-site facilities 14.5Agricultural facilities 11.2Planning and design unit .657Government provides land and plans facility 92.1Government provides land and private investor

designs facility5.9

Private investor provides land and designs facility 2.0PPP strategy .000BOT 49.3OT 32.2ROT 18.4Major public infrastructure .000No 61.2Yes 38.8Project scale (thousand NTD) .000<5,000 15.85,000–50,000 15.8>50,000 68.4PCIR (%) .057<25 3.325–50 0.050–75 3.9>75 92.8LOD (year) .000<5 19.75–10 23.010–15 5.915–20 13.8>20 37.5

3962 J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964

To evaluate the model performance, classification performancemeasure can be obtained as below

Accuracy ¼ tpþ tntpþ fnþ fpþ tn

� �ð10Þ

Precision ¼ tptpþ fp

� �ð11Þ

Sensitivity ¼ tptpþ fn

� �ð12Þ

Table 4Cross-fold modeling performance.

Classification model Accuracy (%) Precision (%)

Proposed modelGASVM 89.30 (1) 94.67 (1)

Baseline modelCART 80.00 85.25 (2)CHAID 82.63 (3) 84.47 (3)QUEST 79.06 81.86C5.0 83.25 (2) 84.24

Previous workEnsemble approach (Chou & Lin, 2013) 84.33 85.60% Improved by GASVM 5.93 10.6

(1)–(3) Denotes performance ranking; The bold values represent the best performance m

Specificity ¼ tnfpþ tn

� �ð13Þ

AUC ¼ 12

tptpþ fn

� �þ tn

fpþ tn

� �� �ð14Þ

Based on the above measures, the following overall average per-formance score (S) is proposed

S ¼ 1m

Xm

i¼1

Pi ð15Þ

where m represents the number of distinct performance measuresand Pi denotes the ith performance measure. The S range is 0–1;the coefficient positively correlates with the effectiveness of theoverall evaluation measures.

5. Results and models comparison

The classification analyses were reproduced by cross-fold meth-od. In each fold experiment, GASVM, CART, CHAID, QUEST, andC5.0 were implemented for model training. The testing fold wasthen used to evaluate the model of each method. The procedurewas then rotated to the next fold until all folds were tested. Thecoincident matrices (rows and columns show actual and predictedresults, respectively) for the individual models were used toquantify model performance in terms of five measures: accuracy(Eq. (10)), precision (Eq. (11)), sensitivity (Eq. (12)), specificity (Eq.(13)), and AUC (Eq. (14)). A synthesis index S (Eq. (15)) was alsoderived for each classification model to represent overallperformance.

Table 4 shows the cross-fold modeling performance. Notably,the IBM SPSS modeler, a highly effective and versatile data analyt-ics workbench, was utilized to develop four classification tech-niques as baseline models. All individual classification modelsachieved at least 80% accuracy except QUEST. The GASVM wasthe most accurate in terms of accuracy value (89.30%), the mostcommon single measure of model performance. In terms of overallperformance measure S, GASVM (0.871) ranked highest followedby C5.0, CART, CHAID, and QUEST. Interestingly, C5.0 was the bestmodel for classifying non-dispute examples (sensitivity = 95.58%)while GASVM performed best at predicting dispute/no-disputeoutcomes (accuracy = 89.30%), measuring classification fidelityfor no-dispute examples (precision = 94.67%) and identifyingdispute cases (specificity = 93.64%). Moreover, GASVM was alsothe best classifier in terms of avoiding false classification(AUC = 0.8394).

Previously, Chou and Lin (2013) used ensemble approach toestimate dispute classification. (Chou & Lin, 2013). Table 4 shows

Sensitivity (%) Specificity (%) AUC S

74.24 93.64 (1) 0.8394 (1) 0.871 (1)

89.30 50.39 (2) 0.6985 (2) 0.750 (2)94.42 (2) 44.01 (3) 0.6920 (3) 0.749 (3)93.36 (3) 33.68 0.6371 0.703

95.58 (1) 42.62 0.6910 0.750 (2)

95.26 48.82 0.7229 0.773– 91.8 16.2 12.7

easures.

J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964 3963

that the proposed GASVM had improved rates by 5.93% (accuracy),10.6% (precision), 91.8% (specificity), 16.2% (AUC), and 12.7% (S). Formost of the performance measures, the study obtained more satis-factory results than did the previous work.

6. Conclusions

This study proposes several classifiers that can be applied whenusing CART, QUEST, C5.0, CHAID, and GASVM (a hybrid approach)to predict dispute propensity. In terms of accuracy, GASVM(89.30%) and C5.0 (83.25%) are the two best classification andregression-based models in predicting project disputes. Moreover,GASVM provides the highest overall performance measurementscore (0.871) considering accuracy, precision, sensitivity, andAUC. Notably, with the exception of GASVM, which was developedby the authors and implemented within a mathematical tool, allmodels are easily executed via open-source or commercial soft-ware. Compared to the baseline models (i.e., C5.0, CHAID, CART,and QUEST) and previous work, GASVM provides 5.89–12.95%higher classification accuracy.

Practitioners must consider the tradeoffs between commercialartificial intelligence tools and a self-developed hybrid inferencemodel. Specifically, one should decide whether the improvementin quantitative accuracy obtained by GASVM is worth the added ef-fort needed to develop a new model compared to using readilyavailable classification and regression-based models with provenease of use, convenience and availability. Advanced research canfocus on developing GASVM expert systems with window or brow-ser interfaces to facilitate use by project managers.

Accurately forecasting dispute propensity provides the proac-tive-warning and decision-support information needed to managepotential disputes effectively and to select appropriate resolutionstrategies before disputes occur. Although the proposed classifica-tion techniques have proven effective for early prediction of dis-pute likelihood in PPP projects involving public infrastructureservices, some classification techniques and their variations werenot evaluated in this study.

Moreover, although an RBF kernel function was used in theSVM-based classification experiments in this study, other kernelparameters and other SVM types may also be optimized usingthe same approach. Similar parameter optimization procedurescan be used to solve prediction or clustering association problemsto decision support system. This study confirmed that the pro-posed hybrid method can assist government agencies in earlywarning of dispute propensity, and thereby reducing the timeand effort needed to prepare a rule set to proactively prevent dis-putes among parties.

References

Arditi, D., Oksay, F. E., & Tokdemir, O. B. (1998). Predicting the outcome ofconstruction litigation using neural networks. Computer-Aided Civil andInfrastructure Engineering, 13, 75–81.

Arditi, D., & Pulket, T. (2005). Predicting the outcome of construction litigation usingboosted decision trees. Journal of Computing in Civil Engineering, 19, 387–393.

Arditi, D., & Pulket, T. (2010). Predicting the outcome of construction litigation usingan integrated artificial intelligence model. Journal of Computing in CivilEngineering, 24, 73–80.

Arditi, D., & Tokdemir, O. B. (1999). Using case-based reasoning to predict theoutcome of construction litigation. Computer-Aided Civil and InfrastructureEngineering, 14, 385–393.

Biggs, D., Ville, B. d., & Suen, E. (1991). A method of choosing multiway partitions forclassification and decision trees. Journal of Applied Statistics, 18, 49–62.

Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification andregression trees. New York: Chapman & Hall/CRC.

Chau, K.-W. (2006). Prediction of construction litigation outcome using a split-stepPSO algorithm. Neural information processing (Vol. 4233, pp. 1101–1107). BerlinHeidelberg: Springer.

Chau, K. (2007). Application of a PSO-based neural network in analysis of outcomesof construction claims. Automation in Construction, 16, 642–646.

Chen, K.-Y. (2007). Forecasting systems reliability based on support vectorregression with genetic algorithms. Reliability Engineering & System Safety, 92,423–432.

Chen, J.-H. (2008). KNN based knowledge-sharing model for severe change orderdisputes in construction. Automation in Construction, 17, 773–779.

Chen, J.-H., & Hsu, S. C. (2007). Hybrid ANN-CBR model for disputed change ordersin construction projects. Automation in Construction, 17, 56–64.

Cheng, M.-Y., Tsai, H.-C., & Chiu, Y.-H. (2009). Fuzzy case-based reasoning for copingwith construction disputes. Expert Systems with Applications, 36, 4106–4113.

Cheung, S.-O. (1999). Critical factors affecting the use of alternative disputeresolution processes in construction. International Journal of ProjectManagement, 17, 189–194.

Cheung, S.-O., Suen, H. C. H., & Lam, T.-I. (2002). Fundamentals of alternative disputeresolution processes in construction. Journal of Construction Engineering andManagement, 128, 409–417.

Cheung, S., Yiu, T., & Chan, H. (2010). Exploring the potential for predicting projectdispute resolution satisfaction using logistic regression. Journal of ConstructionEngineering and Management, 136, 508–517.

Chong, H., Mohamad Zin, R., & Chong, S. (2012). Employing data warehouse forcontract administration: e-Dispute resolution prototype. Journal of ConstructionEngineering and Management.

Chou, J.-S., & Lin, C. (2013). Predicting disputes in public–private partnershipprojects: Classification and ensemble models. Journal of Computing in CivilEngineering, 27, 51–60.

Donis-Díaz, C. A., Muro, A. G., Bello-Pérez, R., & Morales, E. V. (2014). A hybrid modelof genetic algorithm with local search to discover linguistic data summariesfrom creep data. Expert Systems with Applications, 41, 2035–2042.

El-Adaway, I. H., & Kandil, A. A. (2010). Multiagent system for construction disputeresolution (MAS-COR). Journal of Construction Engineering and Management, 136,303–315.

Fei, S.-W., & Zhang, X.-B. (2009). Fault diagnosis of power transformer based onsupport vector machine with genetic algorithm. Expert Systems withApplications, 36, 11352–11357.

Fenn, P., Lowe, D., & Speck, C. (1997). Conflict and dispute in construction.Construction Management and Economics, 15, 513–518.

Ferri, C., Hernández-Orallo, J., & Modroiu, R. (2009). An experimental comparison ofperformance measures for classification. Pattern Recognition Letters, 30, 27–38.

Garg, H., Rani, M., & Sharma, S. P. (2014). An approach for analyzing the reliability ofindustrial systems using soft-computing based technique. Expert Systems withApplications, 41, 489–501.

Goetz, J., & Gibson, G. (2009). Construction litigation, U.S. general servicesadministration, 1980–2004. Journal of Legal Affairs and Dispute Resolution inEngineering and Construction, 1, 40–46.

Goldberg, D. E., Deb, K., & Horn, J. (1991). Don’t worry, be messy. In R. Belew & L.Booker (Eds.), Proceedings of the fourth international conference in geneticalgorithms and their applications (pp. 24–30). USA: San Diego.

Goldberg, D. E., Deb, K., Kargupta, H., & Harik, G. (1993). Rapid, accurateoptimization of difficult problems using fast messy genetic algorithm.Australian Electronics Engineering, 27, 56.

Hajdasz, M. (2014). Flexible management of repetitive construction processes by anintelligent support system. Expert Systems with Applications, 41, 962–973.

Hettiarachchi, D. S., Noman, N., & Iba, H. (2013). Messy genetic algorithm for evolvingmathematical function evaluating variable length gene regulatory networks. InIEEE congress on evolutionary computation (CEC), 2013 (pp. 2154–2161).

Hsiao, F.-Y., Wang, S.-H., Wang, W.-C., Wen, C.-P., & Yu, W.-D. (2012). Neuro-fuzzycost estimation model enhanced by fast messy genetic algorithms forsemiconductor hookup construction. Computer-Aided Civil and InfrastructureEngineering, 27, 764–781.

Hsu, C.-W., Chang, C.-C., & Lin, C.-J. (2003). A practical guide to support vectorclassification. Department of Computer Science, National Taiwan University.

Hsu, C.-W., & Lin, C.-J. (2002). A simple decomposition method for support vectormachines. Machine Learning, 46, 291–314.

Huang, J., Bo, Y., & Wang, H. (2011). Electromechanical equipment state forecastingbased on genetic algorithm – Support vector regression. Expert Systems withApplications, 38, 8399–8402.

Huang, C.-L., & Wang, C.-J. (2006). A GA-based feature selection and parametersoptimization for support vector machines. Expert Systems with Applications, 31,231–240.

IBM. (2010). PASW modeler. In NY, USA: IBM Corporation._Ilhan, _I., & Tezel, G. (2013). A genetic algorithm–support vector machine method

with parameter optimization for selecting the tag SNPs. Journal of BiomedicalInformatics, 46, 328–340.

Ilter, D. (2012). Identification of the relations between dispute factors and disputecategories in construction projects. International Journal of Law in the BuiltEnvironment, 4, 45–59.

Irani, Z., & Kamal, M. M. (2014). Intelligent systems research in the constructionindustry. Expert Systems with Applications, 41, 934–950.

Jiao, Y.-B., & Liu, H.-B. (2011). Application of genetic algorithm-support vectormachine (GA-SVM) for damage identification of bridge. International Journal ofComputational Intelligence and Applications, 10, 383–397.

Jin, X.-H., & Zhang, G. (2011). Modelling optimal risk allocation in PPP projects usingartificial neural networks. International Journal of Project Management, 29,591–603.

Jones, D. (2006). Construction project dispute resolution: options for effectivedispute avoidance and management. Journal of Professional Issues in EngineeringEducation and Practice, 132, 225–235.

3964 J.-S. Chou et al. / Expert Systems with Applications 41 (2014) 3955–3964

Kassab, M., Hegazy, T., & Hipel, K. (2010). Computerized DSS for constructionconflict resolution under uncertainty. Journal of Construction Engineering andManagement, 136, 1249–1257.

Keerthi, S. S., & Lin, C.-J. (2003). Asymptotic behaviors of support vector machineswith gaussian kernel. Neural Computation, 15, 1667–1689.

Khashei, M., Zeinal Hamadani, A., & Bijari, M. (2012). A novel hybrid classificationmodel of artificial neural networks and multiple linear regression models.Expert Systems with Applications, 39, 2606–2620.

Kim, Y. S. (2010). Performance evaluation for classification methods: A comparativesimulation study. Expert Systems with Applications, 37, 2292–2306.

Kim, M.-J., & Kang, D.-K. (2012). Classifiers selection in ensembles using geneticalgorithms for bankruptcy prediction. Expert Systems with Applications, 39,9308–9314.

Kim, H.-J., & Shin, K.-S. (2007). A hybrid approach based on neural networks andgenetic algorithms for detecting temporal patterns in stock markets. AppliedSoft Computing, 7, 569–576.

Kim, G.-H., Yoon, J.-E., An, S.-H., Cho, H.-H., & Kang, K.-I. (2004). Neural networkmodel incorporating a genetic algorithm in estimating construction costs.Building and Environment, 39, 1333–1340.

Knjazew, D. (2002). OmeGA: A competent genetic algorithm for solving permutationand scheduling problems. Boston: Kluwer Academic Publishers.

Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimationand model selection. Proceedings of the 14th international joint conference onartificial intelligence (Vol. 2, pp. 1137–1143). Morgan Kaufmann Publishers Inc..

Lee, M.-C. (2009). Using support vector machine with a hybrid feature selectionmethod to the stock trend prediction. Expert Systems with Applications, 36,10896–10904.

Lee, L. H., Rajkumar, R., Lo, L. H., Wan, C. H., & Isa, D. (2013). Oil and gas pipelinefailure prediction system using long range ultrasonic transducers andEuclidean-support vector machines classification approach. Expert Systemswith Applications, 40, 1925–1934.

Liao, S.-H., Chu, P.-H., & Hsiao, P.-Y. (2012). Data mining techniques andapplications – A decade review from 2000 to 2011. Expert Systems withApplications, 39, 11303–11311.

Lin, Y. (2002). Support vector machines and the Bayes rule in classification. DataMining and Knowledge Discovery, 6, 259–275.

Lin, H. T., & Lin, C. J. (2003). A study on sigmoid kernels for SVM and the training ofnon-PSD kernels by SMO-type methods. In Technical report. National TaiwanUniversity: Department of Computer Science and Information Engineering.

Loh, W.-Y., & Shih, Y.-S. (1997). Split selection methods for classification trees.Statistica Sinica, 7, 815–840.

Mareels, I., & Polderman, J. W. (1996). Adaptive systems: An introduction. Boston,Massachusetts: Birkhauser.

Marzouk, M., El-Mesteckawi, L., & El-Said, M. (2011). Dispute resolution aided toolfor construction projects in Egypt. Journal of Civil Engineering and Management,17, 63–71.

Menassa, C., & Peña Mora, F. (2010). Analysis of dispute review boards applicationin U.S. construction projects from 1975 to 2007. Journal of Management inEngineering, 26, 65–77.

Moosmayer, D. C., Chong, A. Y.-L., Liu, M. J., & Schuppar, B. (2013). A neural networkapproach to predicting price negotiation outcomes in business-to-businesscontexts. Expert Systems with Applications, 40, 3028–3035.

Ngai, E. W. T., Xiu, L., & Chau, D. C. K. (2009). Application of data mining techniquesin customer relationship management: A literature review and classification.Expert Systems with Applications, 36, 2592–2602.

Pai, P.-F., Hung, K.-C., & Lin, K.-P. (2014). Tourism demand forecasting using novelhybrid system. Expert Systems with Applications, 41, 3691–3702.

PCC. (2010). The strategies to the promotion of public–private partnerships (inChinese). In Taipei, Taiwan (R.O.C.): Public Constrction Commission, ExecutiveYuan.

Pulket, T., & Arditi, D. (2009a). Construction litigation prediction system using antcolony optimization. Construction Management and Economics, 27, 241–251.

Pulket, T., & Arditi, D. (2009b). Universal prediction model for constructionlitigation. Journal of Computing in Civil Engineering, 23, 178–187.

Quinlan, J. R. (2007). C5. In (Vol. 2011).Seera, M., & Lim, C. P. (2014). A hybrid intelligent system for medical data classification.

Expert Systems with Applications, 41, 2239–2249.Seifert, B. (2005). International construction dispute adjudication under

international federation of consulting engineers conditions of contract andthe dispute adjudication board. Journal of Professional Issues in EngineeringEducation and Practice, 131, 149–157.

Shmueli, G., Patel, N. R., & Bruce, P. C. (2007). Data mining for business intelligence.New Jersey: John Wiley & Sons Inc.

Sokolova, M., & Lapalme, G. (2009). A systematic analysis of performance measuresfor classification tasks. Information Processing & Management, 45, 427–437.

SPSS (2007). Clementine 12.0 algorithms guide. Chicago, USA: Integral SolutionsLimited.

Tan, P.-N., Steinbach, M., & Kumar, V. (2006). Introduction to data mining. PearsonEducation Inc..

Tang, L., Shen, Q., & Cheng, E. W. L. (2010). A review of studies on public–privatepartnership projects in the construction industry. International Journal of ProjectManagement, 28, 683–694.

Vapnik, V. N. (1995). The nature of statistical learning theory. New York: Springer-Verlag.

WBG. (2011). http://ppi.worldbank.org/resources/ppi_glossary.aspx. In (Vol. AccessedApril 5, 2011): The World Bank Group.

Witten, I. H., & Frank, E. (2005). Data mining: Practical machine learning tools andtechniques. Morgan Kaufmann.

Wu, Z. (2005). Water quality model calibration by means of fast messy geneticalgorithm. In Impacts of global climate change (pp. 1–12).

Wu, C.-H., Tzeng, G.-H., & Lin, R.-H. (2009). A novel hybrid genetic algorithm forkernel function and parameter optimization in support vector regression. ExpertSystems with Applications, 36, 4725–4735.

Zhao, M., Fu, C., Ji, L., Tang, K., & Zhou, M. (2011). Feature selection and parameteroptimization for support vector machines: A new approach based on geneticalgorithm with feature chromosomes. Expert Systems with Applications, 38,5197–5204.