Construction Dispute Potentials: Mechanism versus Empiricism in Artificial Neural Networks

Abstract

The booming development of neural network algorithms has shifted the research focus in the field of construction project management from causal investigation to statistical approximation and hence from mechanistic models to empirical models. This paper took construction dispute avoidance as an example and enabled the best efforts to establish paired mechanistic and empirical models to investigate if the pursuit of a mechanistic understanding of construction disputes should be continued. A Bayesian belief network and multilayer perceptron were used for mechanistic and empirical simulations, respectively. A list of critical dispute factors was first identified from the literature and shortlisted by Pearson’s chi-square tests and Pearson product-moment correlational coefficient tests. The structure of the Bayesian belief network was constructed with logical deduction assisted by a further literature review and Delphi surveys. A structured questionnaire survey was conducted to collect quantitative data for factor shortlisting and model quantification. It was revealed that, being assisted with machine learning techniques, both mechanistic and empirical models achieved an accuracy rate of over 95% under ideal conditions. However, Bayesian belief network models predicted better with fewer constraints due to their advantages in reflecting the formation mechanism of construction disputes, while multilayer perceptron models were more constrained by the inconvenience of sourcing high-quality data as model input. This paper demonstrated that it is still necessary to investigate the formation mechanism of construction disputes further for more efficient avoidance strategies. During the investigation of model construction and comparison, this paper also reflected on the interpretation of statistical threshold and proposed that an arbitrary single cut-off point for statistical tests could potentially eliminate factors that should have been included.

1. Introduction

Machine learning, since its invention, has revolutionised the construction industry, and construction project management is no exemption. The combination of critical success factors and machine learning has enabled new management means for construction projects, but it also has shifted the research focus from causal investigation to statistical approximation and hence from mechanistic models to empirical models. A mechanistic model refers to models established upon the formation mechanism of an event, while an empirical model refers to those established upon correlational strength. The former explains why and how an event occurs, and the latter ignores the causal interrelationships and only simulates the input and output of an event.

Mechanistic understandings were greatly developed in early research. Without algorithms that forcibly approximate the statistical associations between critical dispute factors and dispute potentials, early investigations had to understand the mechanisms. Pinto [1] devised a project implementation profile consisting of 10 critical dispute factors. The associations among factors were primarily investigated. The profile served as a diagnostic instrument for practitioners to evaluate the propensity of a project towards disputes. Later, Mitropoulos and Howell [2] established a process-based model to explain the formation of construction disputes. The authors argued that the sources of transaction costs persisted in construction projects and took project uncertainty, contractual problems, and opportunistic behaviours as three basic issues driving the development of dispute formation. Busby and Hughes [3] investigated the path through which significant errors materialised into disputes. The errors were defined as latent conditions that resided in a system and materialised into disputes via certain paths. The path mainly concerned time or repetition that took the errors to materialise. It was argued that the most frequent pathogen was practice-related, which often arose from good intentions. Love et al. [4,5,6,7] continued Busby and Hughes’s [3] work and proposed some conceptual causal models for construction disputes with the dispute-contributing roles of the errors and paths being defined. Meanwhile, Cheung et al. [8] established a fuzzy sets model for evaluating construction disputes, and Cheung and Yiu [9] and Cheung and Pang [10] adopted fault-tree analyses to present the essential elements for a construction dispute to occur. Based the fault tree models, the likelihood of a construction dispute occurring in a hypothesised case was calculated. The authors indicated the inevitability of contractual disputes in complex construction disputes and advocated a positive strategy for addressing the dispute potentials.

Recent research was mostly based on empirical relationships [11,12,13]. While the development of empirical models was boosted with the advanced algorithms, the pursuit of mechanistic understanding of construction dispute formation seemed to be ceased. Empirical models forcibly approximate the patterns of input and output of construction disputes and ignore the causal associations of input and output factors. Early empirical models were only able to approximate such input and output with linear models, which limited the model accuracy rate. The recent boost in machine learning algorithms allows empirical models to simulate non-linear boundaries with a set of combined linear models and hence improves the accuracy rate. This paper aims to investigate if empirical models aided with machine learning could bypass the necessity of understanding the formation mechanism. Contractual disputes in construction projects are taken as an example. Comparable machine-learning-aided mechanistic models and machine-learning-aided empirical models will be established, and their efficiencies and accuracies will be examined in different scenarios.

2. Theoretical Background

There are two types of associations that can be revealed by data, correlation and causation [14]. However, there is an evidenced lack of causal research in the field of construction project management. Most of the research in this field reveals correlations, and many of those alleged to have a causal focus either identify single points of contributory factors or identify correlations between factors and effects. This view is supported by a literature review conducted by the authors on studies reporting causes or causality in the field of project management from 11 selected journals, i.e., International Journal of Project Management, Project Management Journal, Journal of Building Engineering, Journal of Management in Engineering, Journal of Construction Engineering and Management, Engineering Construction and Architectural Management, Construction Management and Economics, Sustainability, Energies, Buildings, and Journal of Civil Engineering and Management. Keywords cause, causing, causal, causality, and causation in titles and abstracts were searched in the journal databases. The search timeframe was limited to the most recent 10 years, from 2012 to 2022. The scope was expanded into the whole of construction project management research. This process retrieved 76 pieces of research, 11 of which discussed technical issues other than managerial issues and were hence excluded from further review. The literature review results indicated that causal investigations were attempted in various managerial objectives and managerial events. In terms of causal investigations, the retrieved research could be divided into three genres, i.e., identification of single points in the causal chain, identification of pieces of correlational relationships, and genuine causal attempts in project management research. Detailed topics and division of the retrieved 65 papers are available in

Table 1. The division of the alleged causal research in construction project management.

Journals	Single Points of Contributory Factors	Correlation Investigations of Factors and Effects	Causal Investigation
J. Manag. Eng.	Abotaleb et al. (2019); Budayan (2019); Liao et al. (2021); Min et al. (2018); Ren et al. (2013); Safapour and Kermanshachi (2019); Santoso and Soeng (2016); Ye et al. (2015); Zhang et al. (2017) [18,19,20,21,22,23,24,25,26]	[27,28,29,30]
J. Constr. Eng. Manag.	Blomberg et al. (2014); Brockman (2014); Gonzalez et al. (2014); Kadry et al. (2017); Koc and Pelin Gurgun (2021); Lestari et al. (2019); Liu et al. (2021); Rosenfeld (2014); Russell et al. (2014); Wong et al. (2016); Yeganeh et al. (2019); Zhang et al. (2020) [31,32,33,34,35,36,37,38,39,40,41,42]	Aljassmi and Han (2013); Assaad and El-adaway (2021); Jelodar et al. (2022); Khanzadi et al. (2018); Le et al. (2014); Maemura et al. (2018); Abdul Nabi and El-adaway (2022); Seyis et al. (2016); Love et al. (2016) [43,44,45,46,47,48,49,50,51]	Albert et al. [16]
J. Build. Eng.		Carretero-Ayuso et al. [52]
Proj. Manag. J.	Denicol et al. [53]; Yang et al. [54]; Yau and Yang [55]	Yap et al. [56]	Ahiaga-Dagbui et al. [15]; Love et al. [57]
Eng. Constr. Archit. Manag.	Adam et al. [58]; Agyekum-Mensah and Knight [59]; Durdyev [60]; Habibi and Kermanshachi [61]; Karami and Olatunji [62]; Shahsavand et al. [63]; Tong et al. [64]; Viles et al. [65]; Wang et al. [66]	Ekambaram et al. [67]; Ma et al. [68]; Seki et al. [69]; Vilventhan and Kalidindi [70]; Cong et al. [71]
Sustainability	Tahmasebinia and Song [72], Bitamba and An [73]	Rahman et al. [74], Araújo-Rey and Sebastián [75]	Ansari et al. [76]
Energies	Ismaila et al. [77], Afelete and Jung [78], Pall et al. [79]
Buildings	Abdellatif and Alshibani [80], El-Sayegh et al. [81]	Sepasgozar et al. [82]
J. Civ. Eng. Manag.		Cheng et al. [83]; Shen et al. [84]; Tanriverdi et al. [85]	Love et al. [86]
Constr. Manag. Econ.	Behm and Schneller [87]	Russell et al. [88]

. Only six pieces of research contained genuine causal attempts at understanding construction project management issues. Ahiaga-Dagbui et al. [15] reviewed construction cost overrun causation. Albert et al. [16] proposed multiple baseline testing as a method for establishing causal evidence but did not realise this proposal. Albert et al. [16] also proposed multiple baseline testing as a method for establishing causal evidence but did not realise this proposal. The other three pieces all adopted the system dynamic method for model establishment but did not establish any quantitative model that could realise predictive purposes. Cardenas et al. [17] established a Bayesian belief network towards a probabilistic causation model and had the relationships embedded in the model structure tested by chi-square tests. This is the only research revealed by the literature search that contained causal thinking and is ready to be used for prediction.

Meanwhile, the review observed a booming development of correlational research in the field of construction project management, and construction disputes made no exception. Early empirical models were only able to approximate such input and output with linear models, which limited the model accuracy rate [89]. The recent boost in machine learning algorithms allows empirical models to simulate the non-linear boundaries with a set of combined linear models and hence improves the accuracy rate. Chaphalkar et al. [12] constructed a multilayer perceptron neural network model to predict the outcome of construction dispute claims. The models consisted of intrinsic factors such as contract conditions, actual site situations, documents presented, etc. A three-layer multilayer perceptron network was tested to have the highest accuracy, approaching 100%. Similarly, Alaloul et al. [11] established a neural network model to assist the evaluations of construction projects based on coordination factors. Three multilayer feed-forward networks with backpropagation and Elman-propagation algorithms were adopted to train and test the model. The authors adopted the mean square error to manifest the distance between the predicted results and the observed results. The mean square error method indicated a relatively high accuracy, with an average value of 0.0231. Yousefi et al. [13] and Alaloul et al. [11] both utilised artificial neural networks as an alternative to logistic regression. Empirical models forcibly approximated the patterns of input and output of construction disputes and improved the model accuracy substantively. However, the accuracy rates of the empirical models were tested in optimal scenarios, which might not be applicable in real construction projects. The high accuracy rate of the empirical model alone might not be strong evidence of the empirical model being superior to the mechanistic one. Comparison of different algorithms existed in the adjacent research area, such as construction litigation. David Arditi conducted a series of investigations into construction litigation to predict the outcome of court decisions, including Arditi, Oksay [90], Arditi and Tokdemir [91], Arditi and Pulket [92], Pulket and Arditi [93], and Arditi and Pulket [94]. The algorithms compared were mostly empirical ones, including integrated prediction models, artificial neural networks, case-based reasoning, and boosted decision tress, and yet the comparison regarding the fitness, efficiency, and accuracy between empirical and mechanistic algorithms was still unknown. This paper intends to construct two comparable mechanistic models and empirical models to predict construction disputes and tests their accuracies and efficiencies in both optimal and practical scenarios. Based on the test results, this paper is going to answer the question as to whether the pursuit of understanding the construction dispute formation mechanism should be ceased or continued. The prediction of issues such as construction litigation, construction disputes, and construction claims is sometimes arguable as it is believed by a genre of researchers that such issues are far too complicated to be predicted by a model. It is expected in this paper that certain abstraction can be realised to reflect the majority of the issues if not all and consequently the empirical and mechanistic algorithms for prediction can be compared.

3. Methods

3.1. Research Design, Sampling, and Data Collection

A list of critical dispute factors was identified from the literature and shortlisted for model construction. The structure of the mechanistic model was established upon the associations among the shortlisted factors. Different neural network algorithms were used to train the models, Bayesian belief network for mechanistic models and multilayer perceptron for empirical models. Factor shortlisting and model quantification used the same data collected from a structured questionnaire survey. After constructing the models, an interview survey was conducted to understand the potential users’ perceptions of the models. Several Delphi surveys were conducted throughout the research process to verify the factor list, data pre-processing method, and model structures.

Purposive sampling was adopted for the questionnaire survey to guarantee the data quality. The questionnaire survey targeted senior project management personnel from the top 15 construction companies in the UK ranked by avenues [95]. Convenient sampling was used for the interview survey. The model users were not limited to senior management personnel, and hence convenient sampling was adopted to randomly select project management practitioners. Ten experts were selected to form the panel of the Delphi survey, with five from academia and five from industry, all with over 20 years of experience in relevant fields.

3.2. Methods for Model Construction

3.2.1. Mechanism—Bayesian Belief Networks

The Bayesian belief network (BBN) is the only algorithm that was revealed to be able to reflect causal relationships in the literature review [17]. It was relatively mature and had been tested in the field of construction project management research and hence was selected for the case of the mechanistic model in this paper. It is a probabilistic model which uses a list of variables as model nodes and uses a directed acyclic graph to represent conditional dependencies among variables. A BBN model captures the believed mechanism among a set of variables associated with the event of interest. The dependencies might be due to being observed, being directly related to the event, or being indirectly related to the event. Each associated variable forms a node in the BBN model. A BBN requires prior information on the conditional distribution of variables to estimate the posterior information of the event. In the context of dispute prediction, the structure of the acyclic graph represents the formation mechanism of construction disputes, with critical dispute factors forming the model nodes and the directed lines between nodes indicating the dependence. The prior information requires a set of conditional probability distribution tables containing information from previous construction projects, such as the probability distribution of event X_i on the condition that a dispute occurs, with event X_i being each critical dispute factor. The mathematical rule of BBN models follows Equation (1). The software used to assist BBN simulation was Netica [96]. More details regarding modelling and reasoning with BBN are available at Darwiche [97].

$$P\left(X_1, X_2,X_3,\dots ,X_n\right)={{\displaystyle \prod }}_{i=1}^{n}P\left[X_i|Parents\left(X_i\right)\right]$$

(1)

3.2.2. Empiricism—Multilayer Perceptron Networks

Similar to the BNN method, the multilayer perceptron (MLP) neural network is also widely applied in the field of construction project management and is perceived as one of the most efficient algorithms for simulating management issues [98,99,100]. It is a supervised learning algorithm that creates a feed-forward neural network. It consists of three types of layers, the input layer receiving input signals for processing, the output layer performing the task of prediction, and the hidden layer of an arbitrary number between the previous two layers. In an MLP model, the data flow from the input layer to the output layer, and neurons in the model are trained with the backpropagation learning algorithm. The goal of the model is to approximate function G as the activation function. The hidden layer does not represent any practical implication, and the structure of the model does not reflect the mechanism. The calculation embedded between the input and hidden layers and between the hidden and output layers contains input vector x, bias vector b, with weight matrices w, activation functions G, and output vector $\widehat{\mathit{y}}$. A basic linear combination of the weighted input features ${{\displaystyle \sum }}^{ }{w}_i{x}_i$ enters through the step of activation function G and is split by the $\widehat{\mathit{y}}$ values, following Equation (2). For a non-linear classification case, multiple perceptrons could be combined in order to simulate the non-linear boundary. The MLP model simulation was analysed by IBM SPSS 25.

$$\widehat{y}=G\left(w_0{x}_0+w_1{x}_1+\cdots +w_n{x}_n+w_{b}b\right).$$

(2)

3.3. Methods for Factor Shortlisting

3.3.1. Shortlist Factors for BBN—Pearson’s Chi-Square Tests

The factors listed in Table 1 account for those theoretically critical to construction dispute formation. Statistical criticality is still sought for model efficiency. The Bayesian belief network model relies on the mechanism that the status assigned to parent nodes triggers the change of the distribution of the child node. In other words, it requires the conditions of a child node to be dependent on the conditions of its parent nodes. Therefore, Pearson’s chi-square test for independence could be applicable in this case.

Pearson’s chi-square, symbolised by χ², evaluates if two variables are associated with a certain population. The value manifests the difference between the observed frequency (T_o) and the expected frequency if the examined variables are independent (T_E). Each χ² value is be associated with a significance level, termed p-value, which rules the independence of investigated variables. The calculation of χ² follows Equation (3). The value of p is positively associated with the probability of independence. A conventional cut-off point at 0.05 is first adopted to shortlist the factors in Table 1 [101]. To further test the model’s capability in predicting construction disputes with a limited number of input variables, a more stringent point at 0.01 is then applied [101]. The software used to assist Pearson’s chi-square calculation was IBM SPSS 25.

$$\begin{gathered} {\chi }^2=\frac{{\left(T_O{}_1-T_E{}_1\right)}^2}{T_E{}_1}+\frac{{\left(T_O{}_2-T_E{}_2\right)}^2}{T_E{}_2}+\frac{{\left(T_O{}_3-T_E{}_3\right)}^2}{T_E{}_3}+\cdots +\frac{{\left(T_O{}_n-T_E{}_n\right)}^2}{T_E{}_n} \\ ={{\displaystyle \sum }}^{ }\frac{{\left(T_O{}_i-T_E{}_i\right)}^2}{T_E{}_i}. \end{gathered}$$

(3)

3.3.2. Shortlist Factors for MLP—Pearson’s Correlation Coefficients

The final model of MLP in SPSS provides the linear coefficient of two associated nodes. The coefficients can be adapted using backpropagation algorithms, in which the current output and the desired output are used to adjust the weights in the output layer and the weights in the hidden layers [102]. It utilises multiple linear combinations of the weighted input to simulate the non-linear boundary. Therefore, the Pearson product-moment correlation coefficient could be applicable in this case.

The Pearson correlation coefficient, symbolised by r, is a statistical measure of the degree of linear correlation between two sets of variables. For two given sets of variables x and y, the calculation of r follows Equation (4). The value of r takes the range of [−1,1]. The closer the absolute value to 1, the stronger the correlational strength manifested. An estimated threshold for dividing a satisfactory correlational strength, r_s, is determined by the number of samples n, following Equation (5) [103]. The software used to assist the Pearson product-moment correlation efficiency was IBM SPSS 25.

$$r_{xy}=\frac{n\left({{\displaystyle \sum }}_{i=1}^n{x}_i{y}_i\right)-\left({{\displaystyle \sum }}_{i=1}^n{x}_i\right)\left({{\displaystyle \sum }}_{i=1}^n{y}_i\right)}{\sqrt{\left[n{{\displaystyle \sum }}_{i=1}^n{x}_i{}^2-{\left({{\displaystyle \sum }}_{i=1}^n{x}_i\right)}^2\right]\left[n{{\displaystyle \sum }}_{i=1}^n{y}_i{}^2-{\left({{\displaystyle \sum }}_{i=1}^n{y}_i\right)}^2\right]}}.$$

(4)

$$r_s=\frac{2}{\sqrt{n}}.$$

(5)

4. Data Collection and Analyses

4.1. Initial Factor List

The notion of the critical success factor was proposed in the context of general management by Daniel [104] and later refined by Rockart [105]. The critical success factor method indicates areas where key jobs should be carried out exceedingly well to make the business a success. Later, this method was widely applied in many fields, including construction project management [48,106,107,108,109,110]. The fundamental quantitative research in the field of construction project management requires critical success factors for the desired or undesired event to be identified first. The research into factor identification for construction dispute avoidance became saturated not long after its first introduction to construction project management [6]. Similarly, many other researchers have observed certain longitudinal continuity [111,112]. Modified from Wang et al. [112], critical success factors for construction dispute avoidance, termed critical dispute factors in this paper, were categorised into six groups, under which a total of 22 factors were identified. The final list of categories and factors, shown as those listed in

Table 2. Preparation of input data for BBN and MLP.

Categories (Code)	Factors	Code	χ2	DF	p	r
Client service gap (CL)	Client past experience	CL_EXP c	7.094	8	0.526	0.139
	Client top management support	CL_TMS a,c	16.895	8	0.031	0.186
	Client financial strength	CL_FS	6.279	8	0.616	0.102
	Client onsite team commitment	CL_CMT a,c,d	21.120	10	0.020	0.213
	Client consultation	CL_CSTN	14.506	8	0.069	0.112
	Client monitoring and management	CL_MnM a,c	18.762	8	0.016	0.117
Contractor service gap (CTR)	Contractor past experience	CTR_EXP a,c	16.459	8	0.036	0.166
	Contractor top management support	CTR_TMS	9.502	8	0.302	0.101
	Contractor technical strength	CTR_TS a,c	17.899	8	0.022	0.131
	Contractor onsite team commitment	CTR_CMT c	8.402	8	0.395	0.157
	Contractor project manager	CTR_PM	7.805	8	0.453	0.105
	Contractor planning and control	CTR_PnC a,c	18.318	8	0.019	0.168
Project characteristics (PC)	Project size	PC_SIZE	6.625	6	0.357	−0.041
	Project complexity	PC_CXT a	15.904	8	0.044	−0.022
	Project innovation	PC_INNO	7.678	8	0.466	0.008
Extrinsic uncertainties (EU)	Site differing	EU_SITE c	13.125	8	0.108	0.151
	Unexpected weather	EU_WEA c	9.515	8	0.301	0.133
	Economic stability	EU_STB	4.902	8	0.768	0.090
Contractual arrangements (CA)	Plans and Specifications	CA_PSP a,b,c,d	25.207	8	0.001	0.188
	Risk identification and allocation	CA_RIA a,b,c,d	25.044	8	0.002	0.273
	Fairness of obligation	CA_OBL a,b,c,d	24.132	8	0.002	0.235
Interactive process (IP)	Communication	IP_COM a,b,c	32.297	8	0.000	0.304
	Misunderstanding	IP_MUD a,b,c,d	21.659	8	0.006	0.218
	Opportunism	IP_OPP a,b,c,d	38.085	8	0.000	0.305

^{a, b, c}, and ^d indicates if this certain factor is used in Model A, B, C, and D.

, was validated by the Delphi panel of 10 experts, 5 of which had over 20 years of working experience in the construction industry and 5 of which held senior posts in academia.

• Client service gap refers to the services provided by the client onsite project team benchmarked against the project requirements;
• Contractor service gap refers to the services provided by the contractor onsite project team benchmarked against the project requirements;
• Project characteristics describe the inherent nature of the project tasks;
• Extrinsic uncertainties concern the hostile ambient environment where the project was delivered;
• Contractual arrangements elaborate risks embedded in the project and assign the risks to the obligated party;
• Interactive processes describe the actions between the client and contractor after legal binding takes effect.

4.2. Measurement and Categorisation of Dispute Potential

Both the BBN and MLP models must simulate the dispute potential. This paper adopts a similar measurement to Diekmann and Girard [89] and Molenaar et al. [113]), who used the product of the frequency of disputes and the severity of the largest disputes to represent dispute potential. However, the Delphi panel questioned the consistency of respondents’ perception of frequency and suggested using the actual number of disputes instead. Thus, the measurement of dispute potential follows Equation (6).

$$\begin{gathered} Dispute Potential=Number of Disputes\times \\ Severity of the largest dispute. \end{gathered}$$

(6)

The dispute potentials of the surveyed samples ranged from 0 to 35. Considering practicality and efficiency, this range can be categorised into three scales. Initially, the authors scaled the dispute potentials as 0 and 1 for minor, 2 to 5 for moderate, and above 5 (not included) for severe. However, both the BBN and MLP models experienced difficulties in approximating the scale of severe. After several trials, the authors determined the scales to be 0 to 3 for early warning, 4 to 10 for warning, and above 10 (not included) for an emergency. Such categorisation was agreed upon by the Delphi panel.

4.3. Data Collected by the Questionnaire Survey

The authors collected the data for factor shortlisting and model compiling by a structured questionnaire survey. Retrospective project information was collected, requesting the respondents to recall a typical construction project in which he/she once participated and answer the questionnaire based on that very project. Snowballing sampling was adopted for the survey, where the human resource departments of 11 Tier 1 Chinese contractor companies were contacted. It was preferred that the respondents had at least 10 years of practical experience in the construction industry. In the end, the survey collected 317 valid samples. Among the collected samples, 255 samples were answered by experts with over 15 years of experience in the construction industry, representing a rate of 80.4%. In order to maintain a 7:3 partition for model construction and validation, 222 of the samples were used for model construction, and the remaining 95 were used for validation.

Power analyses were conducted to understand the representativeness of the sample. For Pearson chi-square tests, the highest degree of freedom encountered in this research was 16. In order to control the type I error α below 0.05 and the type II error β below 0.01, the minimum sample size was 150. For Pearson’s chi-square tests and Pearson’s correlation test, respectively, a total of 150 and 195 samples were required for the power of 99%. Input and output parameters for the power analyses are available in

Table 3. Power analyses for sample size calculation.

	Input		Output
Pearson’s chi-square test	Parameters	Value	Parameters	Value
	Effect size	0.5	Noncentrality parameter λ	37.5
	α err prob	0.05	Critical χ2	26.3
	Power (1-β err prob)	0.99	Total sample size	150
	Df	16	Actual power	0.9903
Pearson’s correlation test	Tails	Two	Lower critical r	−0.14
	Correlation ρ H1	0.3	Upper critical r	0.14
	α err prob	0.05	Total sample size	195
	Power (1-β err prob)	0.99	Actual power	0.9903
	Correlation ρ H0	0

. Therefore, 222 samples met the power requirements.

4.4. Constructing the BBN Model

For shortlisting the initial factor list, Pearson’s chi-square tests were conducted to investigate the independence between factors and construction disputes. The results, including χ² and its associated p-values, are summarised in Table 1. The structured questionnaire survey sent out 1481 invitations, with 317 valid samples being returned, representing a respondent rate of 21.4%. Respectively, 12 and 6 factors were separated by the cut-off point at 0.05 and 0.01 and were marked with letters a and b at the top right of factor codes for constructing Model A and Model B.

The relationships among the 12 factors were investigated to determine the model structure to establish Model A. To present the logical structure, categories were included for cases where multiple factors from one category had been shortlisted. This paper adopted logical reasoning assisted by a literature review and Delphi surveys. Logical relationships were proposed first and then were kept in the structure if support from the literature and statistical analysis was available. The principles of the relationship identification process included temporality, plausibility, coherence, and strength. These, respectively, concerned the precedence of parent nodes to child nodes in relation to time, the scientific plausibility of the effect of exposure on the outcome, the conformance of the relationship to the known knowledge in the literature, and the strength of the association. Eight threads a to h were gathered to establish the Model A structure, which was later validated by the Delphi panel.

• Client top management support, Client commitment → Client monitoring and management [114];
• Contractor technical strength, Contractor past experience → Contractor planning and control [114];
• Client monitoring and management, Project complexity → Client service gap [2,89];
• Contractor planning and control, Project complexity → Contractor service gap [2,89];
• Client service gap → Plans and specification, Risk identification and allocation, Fairness of obligation → Contractual arrangements [2,4];
• Client service gap, Contractor service gap → Communication [115,116];
• Communication, Contractual arrangements → Misunderstanding, Opportunism → Interactive processes [2,9,117,118];
• Client service gap, Contractor service gap, Interactive processes → Dispute potentials [3,9,10].

Statistical verification was conducted for a cross-check of the shortlisted factors, summarised in

Table 4. Cross-check of independence among factor shortlist.

	CL_ TMS	CL_ CMT	CL_MnM	CL_ SG	CTR_TS	CTR_EXP	CTR_PnC	CTR_SG	PC_ CXT	CA_ PSP	CA_RIA	CA_OBL	CA	IC_ COM	IC_ MUD	IC_ OPP
CL_CMT	0.000	—
CL_MnM	0.000	0.000	—
CL_SG	0.000	0.000	0.000	—
CTR_TS	0.000	0.193	0.000	0.076	—
CTR_EXP	0.620	0.541	0.064	0.040	0.000	—
CTR_PnC	0.065	0.025	0.000	0.612	0.000	0.000	—
CTR_SG	0.063	0.273	0.007	0.000	0.000	0.000	0.000	—
PC_CXT	0.369	0.762	0.167	0.018	0.000	0.546	0.252	0.574	—
CA_PSP	0.000	0.005	0.000	0.000	0.282	0.722	0.145	0.124	0.178	—
CA_RIA	0.000	0.000	0.000	0.000	0.029	0.117	0.000	0.005	0.004	0.000	—
CA_OBL	0.000	0.000	0.000	0.000	0.825	0.269	0.025	0.012	0.349	0.000	0.000	—
CA	0.000	0.000	0.000	0.000	0.045	0.289	0.007	0.003	0.005	0.000	0.000	0.000	—
IC_COM	0.000	0.000	0.000	0.000	0.009	0.043	0.007	0.000	0.481	0.000	0.000	0.000	0.000	—
IC_MUD	0.000	0.000	0.000	0.000	0.432	0.387	0.018	0.003	0.263	0.006	0.010	0.000	0.000	0.000	—
IC_OPP	0.000	0.000	0.041	0.001	0.007	0.016	0.051	0.000	0.641	0.007	0.000	0.000	0.000	0.000	0.000	—
IC	0.000	0.000	0.000	0.000	0.001	0.027	0.000	0.000	0.425	0.000	0.000	0.000	0.000	0.000	0.000	0.000

. The cross-check among the factors statistically verified the significance of the above eight threads at the significance level p = 0.01. Therefore, the threads represented relationships that were both statistically and logically significant and together established the model structure, shown as the structure of Figure 1.

The significance p at 0.01 only separated six factors under two categories for Model B, i.e., three factors under contractual arrangements and three factors under interactive processes. As discussed in Model A, contractual arrangements acted as a breeding platform for improper interactive processes to propagate. This provides the structure of Model B, shown as the structure of Figure 2. Model B was also validated by the Delphi panel.

For Model A and Model B, respectively, after creating the nodes and structure in the Netica interface, distributions of conditional probabilities of each node were input into the tables of conditional probability distribution as required by the model structure. The distribution tables were extracted from the structured questionnaire survey. The partition of training and test cases was set to be 7:3. In total, 222 random samples out of the collected 317 samples were used for model training, while the remaining 95 were set aside for later accuracy tests. The structure of Model A was then subjected to a simulation of the joint, marginal, and conditional distributions based on the 222 training samples. Assisted with the Netica software, Model A was compiled. More details regarding the Netica BBN simulation are available at Nosys [96].

4.5. Constructing the MLP Model

For shortlisting the initial factor list, Pearson’s product-moment correlation coefficients were calculated to investigate the correlational strength of listed factors to construction disputes. The coefficient value r is summarised in Table 1. The total number of samples n equalled 317. Following Equation (5), the threshold of the r value was shown as Equation (7), being 0.112. This separated 16 factors, marked with the letter c at the top right of factor codes for constructing Model C. In order to investigate the capability of MLP models in predicting construction disputes with a further refined list of factors, another MLP model comparable to Model B was established. Considering Model B contained six factors as input, the top six factors with the strongest correlational strengths were identified and marked with the letter d for constructing Model D.

$$r_s=\frac{2}{\sqrt{n}}=\frac{2}{\sqrt{317}}\approx 0.112.$$

(7)

MLP models do not require logical structures; all input factors were put at the input layer parallel to each other. In SPSS, the authors selected the 16 variables as covariates and Dispute Index as dependent variables, set the partition of the dataset to be 70% and 30%, customised the architecture with one hidden layer with an incremental number of units in the hidden layer, and used batch as the type of training. For each simulation trialled, SPSS used one consecutive step with no decrease in error based on the testing sample as the stopping rule. More information regarding SPSS MLP simulation is available at IBM [119]. Model C achieved satisfactory prediction results when there were seven units in the hidden layer. The activation functions of the hidden layer and the output layer were, respectively, hyperbolic tangent and softmax. Parameter estimates are summarised in

Table 5. Parameter estimates for Model C.

Predictor		Hidden Layer							Output Layer
		H (1:1)	H (1:2)	H (1:3)	H (1:4)	H (1:5)	H (1:6)	H (1:7)	[DP = 1]	[DP = 2]	[DP = 3]
Input Layer	(Bias)	0.140	−2.526	−3.258	−0.625	0.971	0.462	−0.189	-	-	-
	CL_EXP	−1.152	−1.652	2.447	1.290	−0.762	1.995	−0.281	-	-	-
	Cl_TMS	0.269	1.651	−0.570	−0.555	0.257	−0.500	−1.537	-	-	-
	CL_CMT	−0.892	0.804	0.918	1.475	0.211	0.037	−1.122	-	-	-
	CL_MnM	−1.637	−4.178	−2.241	1.452	−1.823	−0.265	−0.610	-	-	-
	CTR_EXP	1.278	−0.684	−1.219	3.028	1.300	−0.181	1.199	-	-	-
	CTR_TS	5.288	0.092	−0.545	1.857	2.032	−0.547	2.362	-	-	-
	CTR_CMT	1.187	1.685	−0.508	−0.849	−0.217	0.834	0.110	-	-	-
	CTR_PnC	−0.684	0.952	1.328	−1.053	0.673	−1.975	−1.436	-	-	-
	EU_SITE	−1.075	−0.578	−0.228	1.214	−3.205	−2.476	1.707	-	-	-
	EU_WEA	−0.938	0.898	−0.863	1.659	1.947	2.290	0.991	-	-	-
	CA_PSP	0.228	−0.097	2.661	−3.369	−0.897	−0.150	0.851	-	-	-
	CA_RIA	−0.506	0.138	−0.370	1.758	1.288	−0.754	−0.145	-	-	-
	CA_OBL	0.516	−0.070	−1.036	0.246	2.983	0.025	−1.100	-	-	-
	IP_COM	−0.488	2.370	−0.336	−0.362	−0.179	−0.053	1.466	-	-	-
	IP_MUD	−0.691	1.129	−0.890	0.194	−0.876	−0.305	1.870	-	-	-
	IP_OPP	−1.974	2.904	−1.200	−1.335	1.042	−3.769	−2.017	-	-	-
Hidden Layer	(Bias)	-	-	-	-	-	-	-	1.709	−3.844	2.569
	H (1:1)	-	-	-	-	-	-	-	2.403	−4.142	1.494
	H (1:2)	-	-	-	-	-	-	-	−1.127	−2.496	3.425
	H (1:3)	-	-	-	-	-	-	-	−0.874	−2.231	2.723
	H (1:4)	-	-	-	-	-	-	-	−1.376	−1.107	2.513
	H (1:5)	-	-	-	-	-	-	-	−1.277	3.937	−1.780
	H (1:6)	-	-	-	-	-	-	-	0.778	−2.837	1.651
	H (1:7)	-	-	-	-	-	-	-	−1.543	3.736	−1.329

. The same process was repeated for Model D. However, the accuracy of Model D was substantively reduced compared to that of Model C. After trials were iterated, the highest accuracy rate appeared in the results. The same activation functions with Model C were used by Model D.

Considering the decreased accuracy rate of Model D, the authors conducted further simulations in order to test MLP models. The authors input the random 222 samples and selected the partition to be 100% training and 0 testing. The model, named Model E, achieved an acceptable accuracy rate for training samples with the same settings as Model C and D.

4.6. Accuracy Tests with Independent Samples

For BBN models, data extracted from the remaining 95 samples were input to the established Model A and Model B for accuracy tests. While the samples collected both the situations of the identified factors and the dispute situation, the authors input the factor conditions into the nodes in the model and compared the predicted results with the true project dispute situation. Model A consisted of 18 nodes, with 13 of them being the identified factors, 4 being the factor categories engaged for complementing the formation mechanism, and 1 being the output, dispute potential. For a standard setting, only 13 factors require input, while the other 4 categories can be left blank. The same can be said for Model B. As for MLP models, SPSS conducted an automatic partition and output the prediction results for both training and testing cases. All 16 input data were required to enable the prediction. Specifically, for testing BBN models’ capability in accommodating missing input, random input information was erased in order to simulate situations of missing input data. The number of factors and what factors are left blank apart from the four categories was determined randomly by Microsoft Excel’s Randbetween() function.

The classifications of predicted results benchmarked to the observed dispute index are shown in

Table 6. Classifications of accuracy test results for Models A, B, C, D, and E.

Model	Sample Groups	Observed Dispute Potential	Predicted Dispute Potential			Accuracy Rate
			1	2	3
Model A	Training (70%)	1	130	0	0	100.0%
		2	1	53	0	98.1%
		3	1	2	35	92.1%
		Overall percent	59.5%	24.8%	15.8%	98.3%
	Testing complete input (30%)	1	60	0	0	100.0%
		2	0	22	0	100.0%
		3	2	1	10	76.9%
		Overall percent	65.3%	24.2%	10.5%	97.6%
	Testing incomplete input (30%)	1	59	1	0	98.3%
		2	2	20	0	90.9%
		3	1	3	9	69.2%
		Overall percent	65.3%	25.3%	9.5%	93.7%
Model B	Training (70%)	1	129	4	2	95.6%
		2	5	46	2	86.8%
		3	3	4	27	79.4%
		Overall percent	61.7%	24.3%	14.0%	91.2%
	Testing complete input (30%)	1	53	2	0	96.4%
		2	4	18	1	78.3%
		3	3	1	13	76.5%
		Overall percent	63.2%	22.1%	14.7%	89.4%
	Testing incomplete input (30%)	1	53	2	0	96.4%
		2	5	17	1	73.9%
		3	2	3	12	70.6%
		Overall percent	63.2%	23.2%	13.7%	87.6%
Model C	Training (70%)	1	128	1	1	98.5%
		2	1	50	0	98.0%
		3	4	3	34	82.9%
		Overall percent	59.9%	24.3%	15.8%	95.5%
	Testing (30%)	1	59	1	0	98.3%
		2	1	24	0	96.0%
		3	0	2	8	80.0%
		Overall percent	65.3%	26.3%	9.5%	96.1%
Model D	Training (70%)	1	116	11	0	91.3%
		2	20	24	10	44.4%
		3	13	14	14	34.1%
		Overall percent	67.1%	22.1%	10.8%	74.8%
	Testing (30%)	1	54	9	0	85.7%
		2	6	13	3	59.1%
		3	1	4	5	50.0%
		Overall percent	64.2%	27.4%	8.4%	75.4%
Model E	Training (100%)	1	126	3	2	96.2%
		2	9	41	0	82.0%
		3	8	3	30	73.2%
		Overall percent	64.4%	21.2%	14.4%	89.9%

. It is indicated that BBN models maintain better accuracy with limited input information, while MLP models experience a substantial decrease. Specifically, Model A and B exhibited further margin for missing input data. While the random number of random input variables was artificially left blank, Model A and Model B both showed relatively high accuracy rates. Model E was then used to assist in understanding the accuracy decrease from Model C to Model D. The same model and simulation settings as Model D were applied to Model E. In total, 222 random samples were used for analysis with partition being set to 100% and 0%, meaning all samples were used for model training. Then, Model E showed a relatively high accuracy rate for training samples. However, when the same settings as Model E were applied to other simulations, except changing the partition from 100% versus 0% to 70% versus 30%, the accuracy rate dropped to the same level as Model D.

4.7. Expert Perceptions on Models

The authors conducted an interview survey with the respondents from the structured questionnaire survey. The interviewees were presented with pictures of Model A versus Model B and Model C versus Model D with their properties, applications, and accuracies explained. It was stressed that these models should only be used as assistance to their management instead of as an alternative. The respondents were then asked the following questions:

• Which model would you prefer to use in order to be assisted?
• Why did you select this model?
• Why did you not select the other models?
• How convenient would it be for you to collect the input data for Model A and Model B?

The survey results indicated that out of the 97 samples collected, 56 (57.7%) respondents preferred Model A, 22 (22.7%) respondents preferred Model B, 19 (19.6%) respondents preferred Model C, and no respondent opted for Model D. The main advantage of Model A over Model C lay in its transparent structure and capability in accommodating missing input data. One frequent query received from respondents was related to the meaning of the nodes in the hidden layer in Model C. Some respondents expressed their uncomfortable feeling about the black-box setting of Model C. Some other respondents selected Model C over Model A as Model C was “mysterious”. They considered Model A an abstraction from human perception of the formation mechanism and wanted a different way of thinking as an alternative to their own judgement. As for data convenience, the majority of interviewees agreed that it would be time-consuming for them to collect accurate evaluations for the required input variables, especially when information on the senior level of management or other organisations was involved.

5. Discussion

5.1. General Rules for Dispute Avoidance

Pearson’s chi-square test of independence and Pearson’s product-moment correlation coefficient results indicated general rules for dispute avoidance. They produced two shortlists of factors critical to construction disputes. The two tests agreed that contractual arrangement factors and interactive process factors were the most critical factors. Early research into critical success factors for conflicts and disputes in project management [106,120,121] were very focused on the capabilities of the project participants and the characteristics of the project itself. It was Ashley et al. [122] and Chua et al. [108] identified that the quality of contractual arrangements and the smoothness of interactive processes were critical. Specifically, these two categories of factors were not emphasised until Cheung and Yiu [9] and Cheung and Pang [10] investigated the roles of these two categories in the dispute formation mechanism. In addition to the results in the literature, this paper indicated that the criticalities of contract and interaction were at a higher order of magnitude over those of project and participant factors.

Pearson’s chi-square test of independence and Pearson product-moment correlation coefficient both highlighted the strong association between contract quality and client capabilities, which was consistent with the fact that the client was more empowered in setting and explaining the contractual arrangements [112]. Existing research agreed on the fact that interactive process factors were strongly associated with both client and contractor factors [3,123]. Established upon this fact, this research further indicated that interactive process factors, being governed by contractual arrangements, were hence more closely associated with client factors rather than contractor factors. As the very last step triggered already established issues to materialise into formal construction disputes, the criticality of interactive process factors was unquestioned. The data analysis results indicated that the client played a dominant role in setting and explaining contractual arrangements and also in the interactive process with the contractor. Therefore, from the client’s perspective, this implies that project management focus should be more focused on contract preparation and contractor governance via contractual arrangements. Clarification of the contract clauses must be sought and explained to the contractor. From the contractor’s perspective, dispute avoidance requires the contractor to understand the contractual arrangements with every specification being clarified and obligation being agreed upon and abide by the contracting rules in the interactive process.

The two data analysis results also agreed on the association strength of participant factors to construction disputes. Client factors and contractor factors performed comparably in contributing to construction disputes. Although the strength of direct association with construction disputes is not as strong as those presented by contractual and interactive factors, the cross-check among factors indicated a strong association of participant factors with contractual arrangements and interactive processes. Negligence of participant factors, e.g., if an incapable client onsite team was formed or if an incapable contractor won the bidding, would put the project at risk of a non-compliant contract and rough interactions, which might trigger the formation of construction disputes.

The weakest association was presented by project characteristics and extrinsic uncertainties. Early research often listed such technical factors as critical [6]. However, as technology advanced to lower the technical barrier in the construction industry, such factors became less important in affecting the project management results [81,124]. They were also the most unstable factors revealed during the data preparation and analysis process. The final measurement and categorisation of dispute potential went through several trials. The trial process was explained in Section 4.3. The authors observed that the statistical criticalities of factors under these two categories changed over the different scale settings during the trials. The authors also observed that factors under these two categories even passed the independence or correlation test, the p-values or the r-value were almost all at the edges of the threshold, i.e., p = 0.05 and r = 0.112. This implied that the impact of these factors on construction disputes was close to the threshold. Unlike contractual or interactive factors that were critical to construction disputes in most cases, project characteristics and extrinsic uncertainty factors might exhibit influence on some dispute cases but not on many other cases. This would lead to a question of how to devise avoidance strategies for factors with criticality close to the threshold. In other words, does statistical criticality mean practical importance for dispute avoidance? Are factors without statistical criticality not important? The factor project complexity was shortlisted with scale setting X of dispute potential but was not with a different setting Y. The statistical implication would be that the factor was capable of triggering the change of dispute potential from one scale to another in setting X but not capable of doing that in setting Y, but how should we understand its practical implication to dispute potential rather than some statistical settings of dispute potential? By observing the analysis results throughout the trial process, the authors proposed to apply further judgmental evidence based on the specific situation of the ongoing case of a construction project for the final decision on factors close to the threshold. Any further statistical analysis for general rules, if not specific-case based, could still be teetering on the edge. The evidence could be expert opinions or a predictive model assisting with the case-by-case scenario.

5.2. Case-by-Case Assistance for Dispute Avoidance

Established upon the factor lists screened out by Pearson’s chi-square tests and the Pearson product-moment correlational coefficient analyses, the authors further established comparable models to investigate the functions of empirical and mechanistic models. The Project Management Book of Knowledge pointed out that conflict in construction where heavy assets were involved was inevitable [125]. Each of the five principles proposed by the PMBOK, i.e., avoid, accommodate, reconcile, force, and collaborate, had a different fitting scenario [126]. The models proposed in this paper assist in selecting suitable principles to guide conflict management and dispute resolution in construction projects.

After different and suitable rules were applied to shortlist the initial factors, both the MLP model and BBN model were constructed with factors that were logically and statistically associated with dispute potential. Best efforts were made to construct the BBN model structure to make it representative of the dispute formation mechanism. For model simulation, both models were assisted with software popularly used in research and practice. From the perspective of model establishment, a mechanistic model would require more effort. Both empirical and mechanical models performed well with conditions attached, but the constraints of an empirical model appeared to be more stringent than those of a mechanical model. With initial rules for factor shortlisting applied, the authors identified 13 factors for Model A and 16 factors for Model B. As an alternative to the general dispute avoidance rules summarised from the association strength of factors and dispute potential, both Model A and Model B predicted accurately.

For years, researchers have been establishing predictive models, empirical or mechanistic, to assist dispute prediction in construction project management. However, how should we understand the roles of these predictive models and the functions of the predicted results in practice? The very first predictive model was proposed by Diekmann and Girard [89], a logistic regression model analogous to the cholesterol test. Diekmann and Girard [89] provided an interpretation that a high dispute potential predicted by the model did not guarantee a dispute, and neither did a low potential preclude the possibility of a dispute. Now there are many empirical or mechanistic models for dispute prediction established in the literature but seldom have the interpretation of predictive results been made available for model users. Model A and Model B, established in this paper, both have accuracy rates over 95%. A predicted dispute potential index of 3, being emergency, means there is a 95% chance that the investigated project will experience a severe dispute situation. If the current dispute status perceived by the model user is a warning, there would be a 95% chance of the dispute situation deteriorating to the next level. Under this circumstance, it is suggested that immediate actions be taken to reduce the dispute propensity. The authors believe that a 95% accuracy rate should provide sufficient confidence for model users to be assisted with the predicted results.

5.3. Empiricism versus Mechanism

While Models A and C predicted construction disputes with high accuracy rates, Models B and D performed differently in terms of accuracy. When input information was reduced, both BBN and MLP models predicted less accurately, but the BBN model still predicted with a satisfactory accuracy rate, while the MLP model deteriorated significantly, being 89.4% versus 75.4%. Model C would still be considered useful for advising on the dispute propensity in parallel to practitioners’ own judgemental call. However, Model D, which barely predicted better than a random guess, lost its utility as it was unable to provide insights in addition to practitioners’ judgement.

Unlike the prediction of people’s stature, meteorological conditions, and stress deformation, the formation of construction disputes does not have a definite causal relationship and involves too many impacting factors, each of which may explain a small amount of variation in construction disputes. Under this circumstance, both mechanistic and empirical models experience difficulties in predicting construction disputes. With the absence of definite causation, the mechanistic models in this paper captured one plausible explanation of the formation mechanism of construction disputes and performed relatively stable in terms of accuracy and efficiency. Specifically, Model A and Model B remained relatively accurate with more stringent conditions. A random number of random variables were artificially kept unavailable to the models, and both models still presented relatively high accuracy rates, 93.7% and 87.6%, respectively.

MLP models forcibly simulate the statistical association between the input of a list of critical dispute factors and the output of dispute potential regardless of the logical process of how these factors lead to the disputes. With a sufficient number of factors as input, the MLP models were able to approximate the associations via wider pathways, but this changed with a decreased number of input factors. In the case of Model E, when the model accurately approximated the functions of the hidden layer for the training samples, it failed to achieve a comparable accuracy rate for the testing samples even though the testing and training samples were randomly picked from the same reservoir. This indicated that MLP models would be highly demanding on the statistical similarities of samples when input variables were limited. In other words, it had limited capability in accommodating statistical fluctuations of samples. Any fluctuations in the testing samples from the training samples would make the model unstable. When the formation mechanism of an issue is ignored, an empirical model focuses only on the mathematical patterns of the input data without considering the practical logic of the issues’ origination and development process. This makes empirical models require more stringent conditions in order to simulate the mathematical patterns well, such as the requirement for high similarities of samples. This constraint would be more obvious with limited input information. This implies an empirical model established upon one group of samples could not be readily applicable to other cases.

Recent machine learning techniques have helped to improve the accuracy rate of empirical models in the field of engineering. With advanced algorithms, an empirical model could achieve an accuracy rate as high as 96.1%. This has almost reached the ceiling of the accuracy rate of a predictive model. If an empirical model can reach an accuracy rate this high, would it still be necessary to seek the mechanism of the investigated issue? In other words, will algorithm advancement help research bypass the mechanism for prediction? In order to answer these questions, the restrictions of such high accuracy of empirical models must be considered. The deterioration in prediction accuracy from Model C to Model D indicated an insufficiency of empirical models in prediction. Model E explained such insufficiency was due to the model’s incapability of accommodating fluctuations of samples. With reduced input data, an empirical model, even assisted by advanced algorithms, could not predict well. On the other hand, assisted with proper algorithms, the application restriction of a mechanistic model was loosened. With greatly reduced input information, the accuracy rate of Model B was still comparable to Model A, and both Model A and Model B still had margins for further missing input data. Therefore, the previous two questions can be answered after we figure out this question: would input data for predictive models in the field of construction disputes always be abundant? The interview survey indicated inconvenience in collecting input data for Model A and Model C. Therefore, dispute avoidance would still require assistance from an accurate and compact model, which appears to be satisfied by a mechanistic model rather than an empirical one.

6. Conclusions

6.1. Theoretical and Practical Contributions

This paper established an empirical model and a mechanistic model to compare their efficiency and accuracy. Theoretically, this paper contributed in two aspects. First, for general dispute avoidance extracted from critical dispute factor analysis, this study provided a way of evaluating factors with criticality at the threshold of statistical analysis. The authors suggested a case-based evaluation method, such as a predictive model, to understand the criticality of factors whose criticality was around the cut-off point. More importantly, this paper compared mechanistic models and empirical models in different scenarios. It was concluded that mechanistic models predicted better with fewer constraints due to their advantage in reflecting the formation mechanism of construction disputes. Constraints are mainly related to the inconvenience of sourcing high-quality data for predictive model input. Therefore, it would still be necessary to further the mechanical understanding of construction disputes.

Practically, this paper proposed both general rules and case-based rules for construction dispute avoidance. It was indicated that interactive process factors and contractual arrangement factors were at a higher order of magnitude over other critical dispute factors. Since contractual arrangements act as a breeding platform for improper interactive processes to propagate, dispute avoidance requires clients to be more cautious in preparing and explaining contract clauses. As for case-based rules, this paper established two mechanistic models, i.e., BBN Model A and Model B, and one empirical model, i.e., MLP Model C, for different application scenarios, with accuracy rates at 97.6%, 89.4%, and 96.1%, respectively. Model B is a simplified version of Model A while maintaining a satisfactory accuracy rate. It can be applied in scenarios where limited input information is available. It would be an optimal solution for a rapid check of the dispute potential of an investigated project.

The mechanical investigation of issues such as construction disputes in the field of engineering has been interrupted by the advancement of machine learning. Both mechanistic and empirical models, assisting with machine learning, can achieve high accuracy rates, but mechanistic models can suit application scenarios of insufficient input while maintaining satisfactory accuracy. Because situations of insufficient input might be prevalent in reality, this paper suggested that understanding the formation mechanism of such issues should be continued to overcome such barriers.

6.2. Limitations and Recommendations

This paper discussed the necessity of the continued pursuit of mechanistic understanding in the field of construction project management. Construction dispute avoidance was used as an example to demonstrate that such a necessity should be continued. As a representative issue in the field of construction project management, construction disputes have high frequency and severe implications in construction projects. However, more fields in the construction subject should be investigated to test and widen the application of the conclusion of this paper. In addition, this paper selected the Bayesian belief network and multilayer perceptron as examples of mechanistic and empirical models, respectively. Although they were believed to be representative and successful in their respective field, more algorithms could be further tested in order to observe their performance in different scenarios in terms of efficiency and accuracy.