Evaluating the Performance of Taiwan Airport Renovation Projects: An Application of Multiple Attributes Intelligent Decision Analysis

Abstract

Performance evaluation is a vital tool for measuring whether construction projects meet their established objectives, particularly in complex tasks. It helps identify key areas for improvement and enhances resource allocation efficiency. Through precise performance evaluation, managers can make optimal decisions regarding resource use, ultimately increasing project productivity and overall performance. The objective of this study is to measure the production efficiency of airport renovation projects in Taiwan using data envelopment analysis (DEA) and to apply artificial neural networks (ANN) for predicting project quality. DEA effectively handles scenarios with multiple inputs and outputs, providing relative efficiency comparisons among projects and quantifying the potential for improvement. ANN, on the other hand, can learn nonlinear patterns from the data, allowing for accurate predictions of project quality. As construction projects become more complex, ensuring efficient operation within limited resources becomes increasingly crucial. Traditional performance evaluation methods are inadequate for addressing scenarios involving multiple inputs and outputs; therefore, using DEA and ANN offers a more accurate framework for analysis and prediction. The results of this study indicate that, through the DEA model, five projects achieved an efficiency score of 1, while twelve inefficient projects need to reduce defect frequency by 54.76% and increase the progress and budget implementation efficiency by an average of 10.33% to improve performance. The ANN model achieved a classification accuracy of 94.1% and a mean squared error (MSE) of 0.11 in regression predictions. By adopting a data-driven approach, ANN facilitates intelligent decision making and forecasting throughout the construction process. This study provides construction managers with concrete guidelines for resource allocation and quality prediction, thus enhancing project management effectiveness.

1. Introduction

Globally, approximately USD 100 billion is required annually for the maintenance and repair of concrete facilities [1]. According to the American Society of Civil Engineers (ASCE) 2019 Infrastructure Report Card, the United States has 3304 public airports, with capital needs for terminal buildings exceeding USD 6.6 billion and runway reconstruction demands approaching USD 16.9 billion. Over the next decade, a projected funding shortfall of USD 111 billion is expected to meet the demands of aviation infrastructure, including terminal construction and upgrades, runway maintenance and expansion, capacity enhancement, and other infrastructure improvements [2]. In general, facility projects in developed countries involve renovations, whereas those in developing countries tend to be new constructions. The construction interfaces and complexity of renovation projects considerably exceed those of new construction projects. Taiwan will soon transition to being a developed country, and some of its infrastructure will deteriorate due to overuse; therefore, Taiwan is expected to encounter the renovation problems encountered in developed countries and must invest considerable labor and resources into maintaining its facilities and integrating the construction of facilities. Therefore, the performance of efficient construction projects should be evaluated to ensure that renovation projects achieve high construction quality within their schedules and budgets.

Performance is generally defined as the success of an operation in relation to set goals [3]. Performance includes effectiveness, which refers to the degree to which a designated objective is obtained, and efficiency, which refers to whether production resource inputs have been optimized to achieve the maximum output. Efficiency is calculated as the output-to-input ratio, with a higher ratio indicating higher efficiency. The ratio can be used to measure the efficiency of a unit, organization, or department or as an indicator of project performance. Kagioglou et al. [4] described performance evaluation as the process of determining how successful organizations are in attaining their objectives and strategies. Therefore, performance evaluation involves a set of management methods that measure the degree to which an objective is achieved; this evaluation is essential in management control.

Measuring the performance of construction projects is essential to improving project performance and cost effectiveness. Currently, construction organizations measure their performance according to a set of predefined performance indicators; such indicators are governed by an organization’s ability to maintain necessary competencies for the successful execution of construction projects [5]. Lin et al. [6] argued that identifying key performance indicators (KPIs) is a crucial first step in the development of an appropriate framework for evaluating construction management performance. Performance indicators can facilitate effective judgment of standards and degrees to determine whether the performance of a process is acceptable [7].

Airports are national hubs of transportation and economic activity. Their major facilities include runways, taxiways, terminal buildings, airport aprons, control towers, and connecting roads, all of which are crucial for airport operations. In Taiwan, airports are public transport facilities that are directly operated and managed by public agencies. The quality and efficiency of airport renovations are vital for aircraft operational safety and maintenance. Accordingly, the Public Construction Commission of the Executive Yuan in Taiwan announced a quality upgrade program in 2008 for improving the quality of public works on the basis of five quality indicators—environment, safety, strength, aesthetics, and functionality—and performance evaluations.

This study included 17 construction projects that were similar in scale and complexity and were all homogenous airport construction projects. The renovations included pavement improvements to facilities such as airport runways, taxiways, and aprons (including those at military airports). To ensure construction quality, the inspection team inspected deficiencies, materials, and equipment. Finally, evaluation scores and grades were determined. The construction inspection data in this study were derived from the public construction bidding management system (PCBMS) and included their progress, budget implementation efficiency, defect frequency, inspection scores, and five major quality indicators scores and were used to measure performance.

Lewin and Minton [8] argued that an ideal performance evaluation model should be able to (1) derive comprehensive indicators for evaluating resource usage; (2) process data with different units of measurement as well as process quantitative and qualitative indicators simultaneously; (3) cope with changes in the external environment; (4) process evaluation tasks that involve multiple inputs and outputs; (5) allocate weights objectively; and (6) provide in-depth information on resource usage as a reference for policy formulation. On the basis of these criteria, data envelopment analysis (DEA) was selected in the present study to evaluate the performance of airport construction projects and artificial neural network (ANN) predicts project quality. DEA was selected because it can be used to process multiple inputs and outputs and qualitative and quantitative indicators measured in different units; moreover, weight allocation in DEA is executed through linear programming, thus avoiding inaccuracies engendered by human subjectivity. Additionally, ANN is a method in artificial intelligence that guides a computer to process data in a way inspired by the human brain. It creates an adaptive system that the computer can use to learn and continuously improve from data. This learning process enables ANN to recognize and manage dynamic or nonlinear behavior [9], as well as model complex linear and nonlinear interactions with extremely high predictive accuracy [10]. ANN is commonly applied in predictive models to extract knowledge from the vast amount of information generated during construction management processes, providing support for decision making [11]. It can identify hidden patterns and relationships within the input data, enabling predictions based on these insights. Therefore, ANN can learn and establish nonlinear relationships between inputs and outputs of data, allowing the computer to make intelligent decisions with limited human assistance. In particular, there is general consensus among researchers about the efficiency of ANN in modeling, especially in the areas of performance evaluation and estimation [12].

In recent research on engineering management, although numerous studies have explored performance evaluation methods for construction projects, many focus on single-dimensional efficiency measures (such as schedule or cost), lacking a comprehensive method capable of handling multiple inputs and outputs simultaneously. As construction projects become increasingly complex, existing methods struggle to accurately assess the overall effectiveness of engineering projects, leaving a gap in practical management. Specifically, there is a need for the effective integration of various input–output evaluation metrics and detailed improvement suggestions for underperforming projects. Many traditional evaluation methods fail to fully account for these factors, preventing managers from obtaining accurate performance indicators and resource allocation guidance. The motivation of this study is to fill this gap in the literature by introducing a more comprehensive performance evaluation framework using DEA and ANN methods. DEA is selected for its ability to handle multiple input–output variables simultaneously, providing an objective evaluation that helps identify inefficient projects and areas for improvement. ANN, on the other hand, can learn nonlinear patterns from large datasets through its self-learning process, improving predictive accuracy. The combination of these methods will assist in offering specific improvement suggestions for underperforming projects and predicting the impact of defects on project quality, thus providing more precise decision support for construction management.

In the first stage, this study evaluated the performance (efficiency) of 17 airport renovation projects by applying DEA to objectively determine and compare the scope of improvement. The performance evaluation indicators comprised three inputs (progress, budget implementation efficiency, and defect frequency) and six output items (inspection scores, environment, safety, strength, aesthetics, and functionality). These items constituted relatively objective and comprehensive indicators of resource usage for assessing the production efficiency of construction projects in order to improve performance and construction management. Specifically, a DEA model was used for performance evaluation. The model evaluation results can serve as a reference for efficiency evaluation and improvement implementation. Moreover, the comparison results can enable supervisors to understand the importance of changing inputs and outputs, which may enhance the performance of their ongoing projects. Specifically, the evaluation results can inform decision makers about resource usage and thus provide information for policy making. Through DEA, project managers can achieve a thorough understanding of the indicators of poor performance and subsequently determine effective management strategies and directions for improvement. In the second stage, the relationship between defects and project quality was deduced from ANN inductive data, confirming that defects will have an impact on project quality. The training set comprises a collection of labeled airport renovation project data that can be generalized to unlabeled construction projects. The model was then used to make predictions. These predictions can provide strategies for construction management, which can help to enhance the quality of construction and the performance of the project.

2. Literature Review

Performance evaluation is a continuous process of identifying, measuring, and developing organizational performance by aligning individual performance goals with the organization’s overall objectives and mission [13]. Evaluating the performance of construction organizations is essential for their continuous improvement and long-term survival [14]. A performance evaluation entails assessing the performance of an organization and its members at a specific time. Commonly used assessment methods include ratio analysis, regression analysis, multicriteria assessments, and DEA. Ratio analysis can be used to process only a single input and output; multiple inputs and outputs must be integrated into one input and output variable through a weighting method. However, decisions about weight selection are subjective, and evaluations performed using this method cannot confirm whether resource application is efficient and cannot provide management with a direction or scope for improvement. Regression analysis entails the use of an output and various inputs as dependent and independent variables, respectively; the prediction value obtained from the regression is used to predict the efficiency value. This method can assess variables influencing productivity and rank their importance. However, regression results reveal only the central tendency of all data and cannot determine the ratings of efficient and inefficient units. In addition, regression analysis can evaluate only one output variable and cannot be implemented with multiple outputs. Multicriteria assessment comprises two parts: determining the subjective relative weights of evaluation criteria (typically determined through the analytic hierarchy process [AHP]) and ranking alternatives (using the Technique for Order of Preference by Similarity to Ideal Solution [TOPSIS]). However, multicriteria assessment is often restricted by its limited objectivity and rigor in project weight evaluation and its limited impartiality in artificial scoring methods.

DEA can simultaneously process multiple inputs and outputs, and the units of measurement need not be identical, thereby enabling a wider scope of application. Furthermore, the weight of variables is derived from the data, producing an impartial basis for units evaluated under different conditions. The efficient frontier obtained through this method is a line representing the most favorable conditions for each evaluated unit; therefore, this line can be a benchmark for other units and stimulate inefficient units to understand areas requiring improvement. DEA is a multi-attribute decision analysis technique in which the attributes of each evaluated decision-making unit (DMU) are divided into inputs and outputs to maximize relative efficiency. Production frontiers constitute the basis for efficiency measurement in DEA; a production frontier is obtained through mathematical programming and does not require the specification of a production function in advance; moreover, it enables the estimation of production frontiers for desired inputs and outputs. When factual data of DMUs are compared with a production frontier, the relative efficiency (or lack thereof) of the DMUs can be determined and improvements can thus be proposed for attaining desired objectives.

2.1. Performance Evaluation of DEA and ANN

Barros and Dieke [15] demonstrated that DEA can enable management to benchmark best-practice DMUs by calculating scores that denote their efficiency through linear programming. DEA enables the consideration of multiple inputs and outputs and is widely used by academics to measure the production efficiency and performance of construction firms. For example, El-Mashaleh et al. [16] employed DEA to determine the critical success factors (CSFs) for construction firms and aspects requiring improvement; performance indicators included progress, cost, safety, customer satisfaction, and profit. Although the study used DEA to identify the CSFs for construction companies, it did not specifically account for the differences between various types of construction companies. Companies of different sizes and types (e.g., small contractors versus large international firms) may require different strategies when addressing the same critical success factors. Horta et al. [17] proposed a novel DEA-based approach for assessing the weights of KPIs for construction companies. The primary limitation of this study is that it uses a static model to evaluate weights, without considering dynamic factors that change over time. Market conditions and technological advancements in the construction industry can affect the importance of performance indicators, but these changes were not accounted for. Several studies have compared the efficiency and productivity of construction firms in several regions, including China, Japan, and South Korea [18], Hong Kong [19], Europe [20], and Australia [21]. However, these studies share several limitations. First, the comparisons among China, Japan, and South Korea do not adequately consider the cultural, economic, and policy differences between these countries, which can affect the accuracy of the results. Second, research conducted in Hong Kong fails to fully reflect the region’s unique urban construction conditions, such as the impact of spatial constraints on construction efficiency. European studies lack in-depth analysis of the differences in policies and market environments across various countries, affecting the validity of cross-national comparisons. Finally, research in Australia provides insufficient analysis of specific factors such as geographic location and labor structure, limiting a comprehensive understanding of the influences on efficiency. Overall, these studies exhibit shortcomings in considering external environments and market conditions, which impacts the broader applicability of the results.

For construction professionals, accurately measuring and predicting project performance is highly complex, requiring effective modeling tools and techniques [22]. ANN is a classification and prediction algorithm that finds locally optimal solutions. It is both powerful and practical in addressing complex problems and nonlinear modeling in the field of construction engineering [23]. Through the process of input layer, hidden layer and output layer, the data are trained to produce predictions as output. ANN’s powerful learning capability has been used in predicting the performance of civil engineering projects. Georgy et al. [24] employed neural intelligent systems to predict the performance of construction projects. Neural intelligent systems rely on large amounts of high-quality data to predict construction project performance. However, the collection of construction project data may be incomplete or inaccurate, leading to a decrease in model accuracy. Dissanayake and Fayek [25] developed a model which integrates fuzzy sets, ANN, and genetic algorithm for monitoring project performance. Due to the integration of multiple techniques, this method incurs higher computational costs and faces implementation challenges, which hinder its widespread application in construction projects. Jha and Chockalingam [26] identified key factors using the ANN model to predict the quality performance of construction projects. The key factors identified by the ANN model depend on the quality and scope of the input variables, and an insufficiently comprehensive variable selection may lead to inaccurate predictions. Fanaei et al. [27] used neuro-fuzzy techniques for quantitatively measuring and predicting six KPIs of construction projects. Although neuro-fuzzy techniques are effective for the qualitative measurement and prediction of construction project KPIs, the complexity and opacity of the model make it difficult to adopt widely. Additionally, the model relies on expert knowledge for adjustments, which may introduce bias due to individual expert perspectives. Tiruneh and Fayek [28] proposed a hybrid neuro-fuzzy system (NFS) that combines the evolutionary optimization techniques of genetic algorithms with a multi-output adaptive neuro-fuzzy inference system (MANFIS) to address multi-input, multi-output (MIMO) problems in construction management performance. While the hybrid NFS approach can manage MIMO issues, the optimization process requires significant computational resources, potentially leading to inefficiencies in practical applications. Umuhoza and An [29] developed an ANN model to predict the quality performance of construction projects in Rwanda. However, as this ANN model focuses on Rwandan projects, it may have limitations in regional applicability. Construction environments and conditions vary across regions, which can affect the model’s generalizability to other areas.

2.2. Critical Success Factors and Performance Indicators

Chan et al. [30] analyzed the relevant literature and demonstrated that different researchers have applied different criteria for construction project success. Chan et al. [31] sorted out five main CSFs in the construction field: human-related factors, project-related factors, project procedures, project management actions, and the external environment. Ramlee et al. [32] reviewed articles reporting project success in 11 leading journals and indicated that most researchers considered cost, time, quality, and management to be the CSFs for construction projects. Chua et al. [33] verified that the CSFs for construction projects included budget, progress, and quality. Therefore, if a project is completed as scheduled and the desired performance objectives are achieved within the budget and at the required quality, the project can be considered successful [34]. In addition, the identification of CSFs can provide indicators for project participants, and they can be used for suggesting critical factors requiring attention to ensure project success. In general, each CSF has quantifiable KPIs [6].

Performance indicators are parameters used to quantify the efficiency or effectiveness of past actions [35]. Cox et al. [36] stated that performance indicators can be either quantitative or qualitative. Critical performance indicators can be derived from the conduct of construction workers, such as quality control, timely completion, cost, safety, and productivity; the most widely applicable indicators can be used to evaluate the performance of a construction project. Therefore, KPIs are quantifiable measures used to assess or compare performance relative to strategic and operational goals [37]. Researchers have developed various KPIs to evaluate the performance needs of construction projects. For example, Kometa et al. [38] adopted a comprehensive approach to evaluate the performance of construction projects; the approach entailed defining a set of KPI criteria that included safety, construction costs, operational and maintenance costs, time, and flexibility for users. Hwang and Zhao [39] reviewed seven global performance measurement and benchmarking initiatives to understand their implications for the Singaporean construction industry. Moreover, they established a nationally recognized performance system reflecting Singapore’s KPIs of cost, time, quality, safety, and related goals. Omar and Fayek [5] identified seven performance categories (cost, schedule, change, safety, quality, productivity, and satisfaction) for 46 projects and quantified the relationship between construction project capacity and KPIs. Madushika et al. [40] surveyed 150 professionals in the Sri Lankan construction industry by using questionnaires to determine the most appropriate KPIs that are meaningful at the construction stage.

The conventional approach to evaluating construction project performance is based on factors such as cost, quality, and time [41]. Ofori-Kuragu et al. [3] demonstrated that construction KPIs share common themes, such as time, cost, quality, deficiencies, and productivity. In Taiwan, codes and regulations for construction management and inspections of major construction projects primarily rely on the rate of progress, budget implementation efficiency, and construction quality management as indicators for performance evaluation. This is consistent with the findings of Chua et al. [33] who identified CSFs for different project objectives, including budget, schedule, and quality. Although the CSFs for some projects differ, such factors can help project managers effectively allocate limited resources and achieve high construction performance [42]. Accordingly, the present study reviewed the literature on the following construction project performance indicators: progress, budget, defects, and quality (including environment, safety, strength, aesthetics, and functionality).

Table 1. Literature review for construction project performance indicators.

Year	Authors	Progress/Time	Budget/Cost	Defects	Quality
2001	Kagioglou et al. [4]	●	●	●	●
2003	Cox et al. [36]	●	●	●	●
2004	Bassioni et al. [43]	●	●	●	●
2006	Menches and Hanna [44]	●	●	●	●
2008	Luu et al. [45]	●	●	●	●
2009	Chan [46]	●			●
2009	Skibniewski and Ghosh [47]	●	●		●
2010	Horta et al. [48]				●
2013	Alzahrani and Emsley [49]	●	●	●	●
2013	Langston [50]	●	●		●
2013	Rimbalova and Vilcekova [51]	●			●
2018	Alaloul et al. [12]	●	●		●
2019	Amarkhil and Elwakil [52]	●	●
2020	Madushika et al. [40]	●	●		●
2023	Maya et al. [53]	●	●		●
2023	Aboseif and Hanna [54]	●	●
2024	Homthong et al. [55]	●			●

lists the results of the literature review.

3. Methodology

3.1. Research Framework and Process

This study aims to enhance performance evaluation and quality prediction for airport renovation projects using DEA and ANN. Figure 1 illustrates the main steps of the research, which are detailed as follows:

• Data collection: First, data related to airport renovation projects are collected from the PCBMS. This data includes performance indicators such as progress, budget implementation efficiency, defect frequency, inspection scores, and five quality indicators. The completeness and accuracy of the data are crucial for the subsequent analysis.
• Data envelopment analysis: The second step of this study involves using DEA to measure the production efficiency of 17 airport renovation projects. DEA is a non-parametric performance evaluation technique used to assess efficiency by comparing the inputs and outputs of different projects. In this process, appropriate input and output variables are established, and the efficiency values for each project are calculated.
• Inefficient DMUs: At this stage, the analysis focuses on the identified inefficient projects (decision making units, DMUs) to determine potential efficiency improvement strategies. By calculating the efficiency indicator ratios of the inefficient projects, specific improvement recommendations can be made, along with a quantification of the extent of those improvements.
• ANN architecture design: The fourth step of this study involves designing the architecture of the ANN. In this process, the number of hidden layers, the number of neurons in each layer, and the learning parameters are determined. These design choices will influence the model’s learning capacity and predictive accuracy.
• ANN training and testing: At this stage, the collected data on defects and inspection scores are used to train the ANN. After training is completed, testing is conducted to evaluate the model’s performance and confirm the effectiveness of the ANN in predicting project quality.
• Decision making in construction: In the final step, the relationship between defects and project quality is examined based on the results from DEA and ANN, providing recommendations for decision making in construction management. This analysis helps project managers better understand how to adjust construction plans and resource allocations based on the predictive results, thereby enhancing the overall quality and efficiency of the projects.

As an example, DEA was also applied to the aforementioned data to evaluate the performance of 10 companies; the number of employees, amount of equipment, and operating revenue of the companies were for evaluation. This evaluation was based on two inputs and one output, which were plotted in a two-dimensional coordinate space, with the horizontal axis representing the number of employees/operating revenue (y₁/x) and the vertical axis representing the amount of equipment/operating revenue (y₂/x) (

Table 2. Input and output data of 10 companies.

Company	A	B	C	D	E	F	G	H	I	J
Operating revenue (x)	1	2	2	3	1	2	2	3	1	2
Number of employees (y1)	4	4	6	6	2	8	1	3	1	4
Amount of equipment (y2)	3	6	4	12	2	4	8	6	3	2
y1/x	4	2	3	2	2	4	0.5	1	1	2
y2/x	3	3	2	4	2	2	4	2	3	1

). The plot for each company is illustrated in Figure 2. A company yielding high output (e.g., operating revenue) along with low input (number of employees and equipment) was considered to have excellent business performance and thus favorable efficiency. Therefore, an efficient frontier could be established by connecting companies G, H, and J in Figure 2. These three companies were determined to be efficient because of their small input and large output. The remaining seven companies inside the efficient frontier were determined to be inefficient units.

When the line representing company E (Figure 2) was connected to the origin (0,0), the line segments between the origin and E (OE) and between H and J (HJ) intersected at E′. Accordingly, the efficiency value of company E could be derived as OE′/OE, which was less than 1. Thus, the efficiency values of companies G, H, and J on the efficiency frontier were all 1. For an inefficient company to achieve efficiency, the input value of the abscissa or ordinate (i.e., employees or equipment) must be reduced to produce a coordinate located on the efficiency frontier.

3.2. DEA Process

Charnes et al. [56] proposed a DEA model under the premise of constant returns to scale (CRS); this model is generally referred to as the Charnes–Cooper–Rhodes (CCR) model. It entails the use of linear programming to obtain efficiency values in the form of percentages, thereby circumventing difficulties in item weight allocation and enabling an objective performance evaluation. The input-oriented CCR model establishes a total of N same-industry or ‘homogeneous’ DMUs (which, in the scope of this study, comprised airport construction projects). Each DMU requires an input of m resources to produce s outputs. If DMU j requires an input of x_ij (i = 1, …, m) resources for an output of y_rj (r = 1, …, s) resources, the efficiency of the kth DMU (DMU_k) can be estimated using the input-to-output ratio h_k, which is represented as a percentage. Therefore, 0 ≤ h_k ≤ 1. If u_r is the weight of the rth output and v_i is the weight of the ith input, the relationship among u_r, v_i, and h_k can be expressed as:

$$\begin{gathered} Maximize:h_k={\displaystyle \frac{{\sum }_{r=1}^s{u}_r{y}_{rk}}{{\sum }_{i=1}^m{v}_i{x}_{ik}}}, \\ Subiectto:{\displaystyle \frac{{\sum }_{r=1}^s{u}_r{y}_{rj}}{{\sum }_{i=1}^m{v}_i{x}_{ij}}}\le 1(j=1,\cdots ,N) \end{gathered}$$

(1)

where ε represents the non-Archimedean constant, which is an extremely small positive number at 10⁻⁴. The first step is to transform Equation (1) from fractional programming to linear programming by assuming that the denominator (${\sum }_{i=1}^m{v}_i{x}_{ik}$) equals 1, which is subsequently added as a constraint. Multiplying the two ends of Equation (1) by ${\sum }_{i=1}^m{v}_i{x}_{ij}$ yields:

$$\begin{gathered} Maximize:h_k={\sum }_{r=1}^s{u}_r{y}_{rk}, \\ Subiectto:{\sum }_{r=1}^s{u}_r{y}_{rj}-{\sum }_{i=1}^m{v}_i{x}_{ij}\le 0(j=1,\cdots ,N), \\ {\sum }_{i=1}^m{v}_i{x}_{ik}=1, \end{gathered}$$

(2)

The constraint for Equation (1) is that the ratio of ‘actual output’ and ‘actual input’ for each DMU is a value between 0 and 1. The optimized values of u_r and v_i are not predetermined by project managers but obtained from the estimated DMU efficiency values in Equation (1). When h_k = 1, the DMU is deemed ‘efficient’; conversely, when h_k < 1, the DMU is deemed ‘inefficient’. This technique of comparing relative efficiency is fair and objective because every DMU and the corresponding input and output in Equation (1) must act as objective functions, and the inputs and outputs of other DMUs must act as constraints.

To date, a wide range of analytical DEA models have been designed for different situations. The selection of a DEA model involves two considerations: whether to be input- or output-oriented, and whether to emphasize constant or variable returns. From these considerations, four basic DEA models can be derived. Input and output orientations primarily concern whether the DMUs have control over input or output factors, depending on the variables involved. An input orientation involves the pursuit of minimal input while maintaining current output; an output orientation entails pursuing maximum output at the current input level.

Constant–variable returns refer to the relationship between input and output factors, which can be constant, incremental, or decremental. If increasing the resource input by 1% increases the output by 1%, the relationship involves CRS. However, if increasing the input by 1% increases or reduces output by more than 1%, the relationship involves variable returns to scale (VRS). Increasing the output by more than 1% indicates increasing returns to scale; otherwise, the relationship involves decreasing returns to scale. When comparing DMUs, the CRS model assumes that each DMU has CRS. By contrast, the VRS model uses three return conditions (constant, increasing, and decreasing). Therefore, CRS modelling entails comparing all DMUs by using the same conditions; the conditions of VRS modelling can vary by individual DMU. This indicates that CRS models are more rigorous than other models.

3.3. Input and Output Variables for DEA

The DEA model can compute multiple airport inputs and outputs within a single analysis, the model also requires less information for analysis than do other models [57]. Lishan et al. [58] proposed that multiple DEA production inputs and outputs can be optimized using linear programming to evaluate the relative performance of an organization. Currently, DEA is a prominent nonparametric technique for performance evaluation.

DEA primarily focuses on the selection of input and output variables for use as performance indicators. This study evaluated the efficiency of airport construction projects from the perspective of construction project managers. Data on 17 airport construction projects (i.e., DMUs) were collected from the PCBMS. The evaluation indicators were progress, budget implementation efficiency, defect frequency, inspection scores, environment, safety, strength, aesthetics, and functionality. These indicators have been formulated by experts and scholars commissioned by the government and are mainly based on Taiwan’s construction characteristics and experience. They have been continually revised during implementation to resemble actual operations, and they can be used to assess construction quality. In addition, when performing regular construction inspections, the competent government authority in engineering invites three experts with engineering backgrounds to form an inspection team and assess the quality of the construction site. All inspection results and data (that include the five major indicators) are recorded in the PCBMS. For evaluation results to be meaningful, performance evaluators must formulate performance indicators according to the characteristics of the evaluated target. In this study, relevant quantifiable performance indicators were used to measure construction project quality. The indicators are described in detail as follows (

Table 3. Explanation of performance indicators and equations.

Performance Index	Performance Index Description	Equation
Progress	Percentage of actual and scheduled progress on the day of construction inspection	$1-({\displaystyle \frac{\mathrm{Planned}\mathrm{progress}}{\mathrm{Actual}\mathrm{progress}}}\times 100\%)$
Budget implementation efficiency	Percentage of expended and planned budget on the day of construction inspection	${\displaystyle \frac{\mathrm{Expended}\mathrm{budget}}{\mathrm{Planned}\mathrm{budget}}}\times 100\%$
Defect frequency	Total number of defects identified by the inspection team	${\displaystyle \frac{1}{\mathrm{Number}\mathrm{of}\mathrm{defects}}}$
Inspection scores	Average total score rated by the inspection team	A+ > 90; A: 80–89; B: 70–80; C < 70
Environment	Cleanliness of and amounts of dust, noise, and vibrations in the construction site and surrounding areas	Sum of environmental scores
Safety	Measures to prevent falling (e.g., handrails and covered openings), building collapse, electrical leakage, and intrusion (e.g., warning signs, fences, and debris nets) as well as traffic safety measures and applications for hazard notices for the working site	Sum of safety scores
Strength	Quality control measures such as the inspection and testing of concrete, steel bars, formwork, earthworks, structures, and other materials	Sum of strength scores
Aesthetics	Overall aesthetic impression of the site and surrounding areas, including structures, buildings, interior designs, and other aspects	Sum of aesthetics scores
Functionality	The degree to which the outcome meets the demands of the owner as well as its budgetary and economic efficiency	Sum of functionality scores

):

3.3.1. Progress (x₁)

Cooke-Davies [59] discovered that schedule delays are correlated with increased project costs. Therefore, construction projects that can be completed on time within the schedule specified by contracts in which each stage of the execution process is on schedule or the progress is ahead of schedule are deemed to have satisfactory performance while maintaining project costs. Progress is recorded as the percentage of actual progress to that of planned progress. When an inspector determines that progress is 5% ahead of schedule, it is recorded as +5%; conversely, if progress is 5% behind, it is recorded as −5%. However, because the input variables for DEA must be positive, progress in the present study must be converted into the rate of completion; for example, −5% progress was noted as 95% completion.

3.3.2. Budget Implementation Efficiency (x₂)

The rate of budget implementation indicates whether the project is on schedule. A low rate suggests a postponement, delay, or deviation in the schedule that affects the timing of budget expenditure. This also implies that the project may have failed part of the construction acceptance or material inspection and that the project payment cannot be fully provided. Therefore, budget implementation efficiency is a critical indicator of construction project efficiency. In this study, budget implementation efficiency was calculated as the percentage of the expended budget relative to the planned budget. Project owners settle payments on the day of the construction inspection and provide them to the inspection team as an evaluation reference.

3.3.3. Defect Frequency (x₃)

Defect frequency is a key indicator of construction quality as well as the performance of construction project management. An inspection team investigates circumstances at the construction site and records defects on inspection forms. The team then requests that the contractor make improvements within a time limit. The total number of defects is the defect frequency. Generally, a lower defect frequency indicates higher construction quality, and a higher defect frequency indicates lower management performance. Input and output are positively correlated in DEA models, meaning that a 1% additional input leads to ≤1%, rather than −1%, additional output. Therefore, the defect frequencies in this study must be converted into reciprocal forms to conform to the positive correlation in the DEA model.

3.3.4. Inspection Scores (y₁)

High construction inspection scores typically indicate high quality control performance, high construction quality, adequate progress management, and favorable project design. Specifically, it reflects favorable overall performance. Therefore, the score reflects the efficiency (or lack thereof) of contract execution management. The construction inspection score is calculated as the mean of the individual inspector scores. A score of >90 points represents an A+ rating, that of 80–90 points represents an A rating, that of 70–80 points represents a B rating, and that of <70 points represents a C rating. The inspection team discusses the quality of each project and determines the number of points deducted according to the severity of the defects. The total number of points deducted can affect the score. According to a previous study conducted by the current author [60], defect frequency, and score have a weak negative correlation (r = −0.3). When correlation coefficients of the two variables are excessively high, the capacity of a DEA model to distinguish the relative efficiency of the DMUs is reduced, and achieving effective linear programming solutions is impractical. In this situation, only one variable can be selected.

3.3.5. Five Quality Indicators

Indicator scores were based on five major indicators: environment, safety, strength, aesthetics, and functionality. These comprise key items for public project inspectors in assessing the quality and performance of a project. Each quality indicator had a maximum score of 10 points, with 9 to 10 points indicating qualified, 8 to 9 points indicating slight defects, 5 to 8 points indicating general defects, and <5 points indicating severe defects. Summing the scores for all items of each quality indicator yielded a set of scores. The five indicators are defined as follows:

• Environment (y₂)
  • Implementing environmental monitoring, preventing noise generation, and avoiding effects on environmental quality.
  • Implementing on-site environmental protection to reduce air pollution and noise during construction.
  • Maintaining cleanliness inside the construction site and in the surrounding environment as well as a dust-free site entrance (vehicles entering and exiting the site are washed).
• Safety (y₃)
  • Preventing occupational disasters and ensuring work safety and health.
  • Ensuring fall prevention for workers.
  • Preventing building collapses.
  • Preventing electrocution for construction workers.
  • Installing construction warning facilities and external protective nets.
  • Establishing traffic safety measures to ensure pedestrian safety.
• Strength (y₄)
  • Quality that meets design, drawing, and contract requirements.
  • Correct implementation of inspection and control of materials and equipment.
• Aesthetics (y₅)
  Ensuring general aesthetic coordination of buildings, structures, decoration, miscellaneous projects, and landscaping inside and outside the construction site with the surrounding environment.
• Functionality (y₆)
  • Developing a design that meets the engineering objectives.
  • Developing a design that conforms to constructability and maintainability.

3.4. Artificial Neural Network

ANNs have a high capacity for learning and often exhibit exceptional prediction performance on high-dimensional or nonlinear problems that are challenging to formulate with clear mathematical models. Their advantage is in their ability to model complex abstractions in the data via an architecture composed of multiple nonlinear learning layers. Each layer corresponds to unique nonlinear abstractions, resulting in deep networks having a higher representational power [61]. ANNs learn a model, either for classification or regression, from labeled training data by constructing a nonlinear model of variables. The model inputs certain features of a sample and produces an output of the corresponding label for that sample, it can be then used to make predictions on unknown samples.

During the learning process of ANN, the weighted sum of each neuron is calculated using inputs (x_i), weights (w_i), and biases (b_i) as shown in Equation (3). Neurons gain information from the training data and through the network structure and different learning methods, the ANN is trained to produce expected outputs. ANN randomly initializes input parameters during training and iteratively calculates the output results, comparing them with the actual results to obtain the loss function. The parameters are then updated until the loss function reaches a minimum. The loop stops when the error threshold is met. The loss function value measures the discrepancy between the output and predicted values, and it serves as feedback to adjust the weights and reduce the error value. The weights and biases of the neurons are updated according to the error value in order to minimize the loss function value, which measures the training error of the samples, as shown in Equation (4). Where Yi’ and Yi are, respectively, the true label and the predicted probability distribution. Therefore, this study utilizes the defects and inspection scores from PCBMS projects to train and test the ANN model. Finally, the performance of the prediction model is evaluated to confirm the impact of defects on project quality.

$$y_i={\sum }_i^n{w_i·x}_i+b_i,$$

(3)

$$Loss=-\sum Y_i^{\prime }·\mathrm{l}\mathrm{o}\mathrm{g}\left(Y_i\right),$$

(4)

4. Analysis and Discussion

DEA is a nonparametric technique, meaning that it can rely entirely on available data for analysis without a predetermined quantitative parametric model. DEA can assess the performance of multiple variables simultaneously.

4.1. Analysis of DEA Results

When DEA is used to execute a performance evaluation for multiple attributes, the input and output variables are determined, and homogenous DMUs are generated as evaluation and comparison objects. DMUs are examined for their efficiency, and reasons for DMU inefficiency are identified for the proposition of corrective actions. The purposes of input- oriented and output-oriented calculations differ; therefore, the performance values derived from such calculations also differ, but the difference between effectiveness and ineffectiveness is consistent. When DEA is applied, both the purpose of the evaluation and the minimum input required in the current output situation should be determined. Considering the uncontrollability of output variables, whether an airport project uses minimal resources to maintain output quality must be determined. Therefore, this study adopted an input-oriented CCR model for performance evaluation.

Managers are responsible for reducing input, which is much easier to control and execute than increasing project output (score). Therefore, the present study adopted an input-oriented CCR model to evaluate the performance (overall efficiency) of each DMU in terms of progress, budget implementation efficiency, defect frequency, inspection scores, and quality as well as nine input and output variables (

Table 4. Descriptive statistics of input and output variables.

DMUs	Efficiency (Overall Efficiency)	Input			Output (Point)
		Progress (Completion Rate) (%)	Budget Implementation Efficiency (%)	Defect Frequency (Times)	Inspection Scores	Environment	Safety	Strength	Aesthetics	Functionality
		(x1)	(x2)	(x3)	(y1)	(y2)	(y3)	(y4)	(y5)	(y6)
DMU 1	0.97	95.06	100	0.13	83	82	80	86	84	85
DMU 2	0.94	104.53	70.51	0.07	82	86	76	85	85	86
DMU 3	1.00	100	100	0.05	82	86	71	85	87	93
DMU 4	1.00	89.92	100	0.04	81	83	82	85	84	84
DMU 5	1.00	102.76	81.6	0.03	82	81	80	81	82	82
DMU 6	0.98	101.27	100	0.14	88	88	87	89	89	89
DMU 7	0.95	102.66	100	0.06	87	87	85	87	87	87
DMU 8	1.00	98.44	0	0.07	78	79	84	79	79	79
DMU 9	0.94	98.45	100	0.06	83	83	82	85	83	84
DMU 10	0.82	113.74	100	0.11	82	82	80	79	82	82
DMU 11	0.94	100.19	100	0.08	84	86	82	84	86	85
DMU 12	0.91	102	100	0.09	82	83	84	84	82	83
DMU 13	0.89	102.64	100	0.08	81	81	75	80	82	82
DMU 14	0.92	102.95	100	0.11	84	84	80	84	84	84
DMU 15	1.00	101.77	100	0.03	83	83	83	82	83	83
DMU 16	0.94	100.61	96.34	0.09	84	85	83	85	85	85
DMU 17	0.96	104.09	93.09	0.33	88	89	87	89	89	89

). Through fractional programming, the CCR model presented in Equation (1) can be derived. To determine the solution, the denominator of h_k (${\sum }_{i=1}^m{v}_i{x}_{ik}=1$) can be set to 1 for conversion into the linear programming model, as expressed in Equation (2). This model can then determine the maximum efficiency value for the same total input (denominator = 1). Consider, for example, the efficiency value (h₁) of DMU 1: (1) The object function could be used to maximize the output of h₁. Therefore, $h$_i$={\sum }_{r=1}^6{u_{r}y}_{r1}$. (2) This example involved three restrictions. First, the sum of the three inputs of DMU 1 was 1 (i.e., ${\sum }_{i=1}^3{v_{i}x}_{i1}=1$). The sum of the input weights (v_i) could then be subtracted from the sum of the output weights (u_r) of the 17 airport renovation projects and must be ≤ 0, that is, ${\sum }_{r=1}^6{u_{r}y}_{rj}-{\sum }_{r=1}^3{v_{i}x}_{ij}\le 0$ (j = 1,…17) Finally, the weights of the inputs and outputs (u_r, v_i) must be ≥ɛ (non-Archimedean constant), that is, u_r, v_i ≥ 10⁻⁴. In this fractional programming solution, an efficiency value of 0.97 was obtained for DMU 1. The efficiency values of the remaining 16 DMUs were calculated in the same manner. Five airport renovation projects had an overall efficiency of 1: DMU 3, DMU 4, DMU 5, DMU 8, and DMU 15. These DMUs were relatively efficient. The efficiency values of the remaining 12 DMUs were less than 1, indicating potential for improvement.

The progress of DMU 1, DMU 4, DMU 8, and DMU 9 was delayed because their completion rates had not reached 100%; the remaining DMUs had completion rates exceeding 100%. Budget implementation had not reached 100% for DMU 2, DMU 5, DMU 8, DMU 16, or DMU 17. The mean (standard deviation) value of defect frequency was 0.09 (0.07). The defect frequencies observed for DMU 1, DMU 6, DMU 10, DMU 14, and DMU 17 were greater than the mean defect frequency. The inspection scores for the items of the five quality indicators ranged from 71 to 93 points; 75% of DMUs had scores of 75–85 points.

By using the efficiency indicators, airport construction project managers can compare their projects with other high-performing construction projects through DEA linear programming; this can enable them to assess the efficiency of their projects and to improve project performance on the basis of the model output (i.e., efficiency values). Projects with an efficiency value of 1 can be considered as having the highest performance, and this study identified five such projects. These five DMUs, which formed an effective production frontier, were DMU 3, DMU 4, DMU 5, DMU 8, and DMU 15. The remaining projects were deemed ‘inefficient’, with DMU 10 having the lowest performance. Specifically, these construction projects required improvement. A total of 10 DMUs, namely DMU 1, DMU 2, DMU 6, DMU 7, DMU 9, DMU 11, DMU 12, DMU 14, DMU 16, and DMU 17, had efficiency values of 0.9–1; they were thus considered to be ‘marginally efficient units’. After investment reduction, these DMUs had opportunities to become more efficient. However, two DMUs, namely DMU 10 and DMU 13, had efficiency values of <0.9 and were considered to be ‘distinctly inefficient units’ that require relatively large improvements and scales. Efficiency was derived as the input-to-output ratio. Fewer inputs and more outputs were considered to indicate greater efficiency. DMU 10 and DMU 13 were distinctly inefficient; they had relatively high inputs, but the quality indicator scores of their outputs were low. Problems with current construction management models were determined using specific performance indicators. DEA was used to consider the weight of the inputs and outputs of all DMUs, objectively measure relatively inefficient DMUs, and provide improved data for various performance indicators. Accordingly, managers should review the causes of poor internal performance to effectively adjust resource allocation.

Figure 3 presents the proportions of improvements required for the 9 efficiency indicators for the 12 inefficient projects. It also provides goals for potential improvements in the production efficiency of the 12 DMUs. The highest proportion was observed for defect frequency (54.76%), followed by those for progress and budget implementation efficiency (10.33% each).

Among the five quality indicators, the highest proportion of improvement was observed for safety (8.09%), suggesting that construction safety is often ignored during project implementation; moreover, safety measures and worker protection in operating environments constituted the basis of project performance. The second highest proportion was observed for strength (5.81%), indicating that material inspection is a crucial area for ensuring construction quality and performance management. Generally, higher inspection scores indicate higher construction quality and performance in construction management. The findings of the present study verify this, indicating that the 12 inefficient projects required only a 0.42% improvement in the ‘inspection scores’ indicator to become efficient.

DEA can be used to determine the relative efficiency values of different DMUs. In a low-efficiency scheme, inputs should be reduced or outputs should be increased; DEA can identify specific improvement directions. On the basis of changes in the proportions observed for the efficiency indicators (Figure 4), the amount of improvement in inputs and outputs required to make the inefficient projects efficient was determined. For example, to make DMU 10 efficient, the defect frequency must be reduced by 22.10%, and both the progress and budget implementation efficiency must be increased by 17.90%; regarding its outputs, functionality, and aesthetics must be improved by 3.25%, strength by 8.21%, safety by 5.02%, and environment by 2.25%. Diagrams that illustrate such information are instrumental in helping project managers to rapidly identify aspects requiring improvement and reallocate resources for optimized efficiency.

The slack variable model, which uses slack variable-weighted averages and distances, was developed by Tone [62]. This model uses a non-radial estimation method that considers both input and output slack. The estimated efficiency is between 0 and 1. When the efficiency of a DMU is 1, no slacks are in its input or output on the production boundary. The slack variable represents the degree to which the input must be reduced or output must be increased such that a relatively inefficient DMU becomes efficient; it is thus an indication of the required change for improvement. Therefore, managers can use slack variable analysis to understand the input resources of inefficient DMUs, identify the sources of inefficiency, and determine the degree to which the performance indicators must improve. Such analysis can enable managers to determine the scope and size of improvement. For a relatively inefficient DMU, particularly that with a projection point on an efficiency frontier formed by input–output variables, the target of efficiency improvement and its difference can be obtained to identify the necessary direction for efficiency improvement.

The present study used DEA slack variables to analyze the proportion of variables with potential for improved performance. For example, DMU 13 was required to reduce defect frequency by 0.03 and increase inspection scores, safety, strength, aesthetics, and functionality by 1.78−7.58. The output observed for environment was 0, signifying that the environment reached target levels in terms of cleanliness, dust, noise, and vibrations (

Table 5. Slack variables for inefficient DMUs.

Defect Frequency (Times)	Inspection Scores	Environment	Safety	Strength	Aesthetics	Functionality
DMU 1	−0.08	0.00	3.00	4.28	0.95	1.98
DMU 2	−0.01	2.28	0.00	11.30	2.33	1.66
DMU 6	−0.09	0.00	2.07	2.65	3.02	2.05
DMU 7	−0.01	0.00	2.03	3.70	3.94	2.99
DMU 9	−0.01	0.00	1.97	2.44	1.87	2.92
DMU 10	−0.04	0.00	1.84	4.01	6.48	2.66
DMU 11	−0.03	0.02	0.00	3.55	3.89	0.94
DMU 12	−0.04	0.42	1.35	0.00	2.17	3.26
DMU 13	−0.03	0.00	1.89	7.58	4.67	1.78
DMU 14	−0.06	0.00	1.96	5.65	3.80	2.88
DMU 16	−0.04	0.00	0.95	2.71	2.77	1.86
DMU 17	−0.27	0.00	0.99	3.09	2.78	1.88
Total	−0.71	2.72	18.05	50.96	38.67	26.86

). In addition, an analysis of the slack variables revealed that for the relatively inefficient projects to reach the most productive scale size, they must reduce their defect frequency (input) by a total of 0.71 and improve their inspection scores, environment, safety, strength, aesthetics, and functionality (output) by 2.72, 18.05, 50.96, 38.67, 26.86, and 23.86, respectively. Safety was the lowest scoring indicator for the 12 inefficient projects, suggesting that safety measures at the construction site were inadequate and that labor safety management should be improved. In addition, strength required improvement, indicating that the construction quality, material, and equipment did not meet specifications.

The objective of optimization (

Table 6. Optimization analysis for inefficient DMUs.

DMUs	Progress (Completion Rate) (%)	Budget Implementation Efficiency (%)	Defect Frequency (Times)	Inspection Scores	Environment	Safety	Strength	Aesthetics	Functionality
DMU 1	92.74	97.56	0.04	83.00	85.00	84.28	86.95	85.98	85.98
DMU 2	98.22	66.25	0.05	84.28	86.00	87.30	87.33	86.66	86.66
DMU 6	99.03	97.79	0.05	88.00	90.07	89.65	92.02	91.05	91.05
DMU 7	98.05	95.51	0.05	87.00	89.03	88.70	90.94	89.99	89.99
DMU 9	93.11	94.58	0.04	83.00	84.97	84.44	86.87	85.92	85.92
DMU 10	93.38	82.10	0.05	82.00	83.84	84.01	85.48	84.66	84.66
DMU 11	94.43	94.26	0.05	84.02	86.00	85.55	87.89	86.94	86.94
DMU 12	92.82	91.00	0.04	82.42	84.35	84.00	86.17	85.26	85.26
DMU 13	91.28	88.93	0.04	81.00	82.89	82.58	84.67	83.78	83.78
DMU 14	94.69	91.98	0.05	84.00	85.96	85.65	87.8	86.88	86.88
DMU 16	94.84	90.81	0.05	84.00	85.95	85.71	87.77	86.86	86.86
DMU 17	100.05	89.48	0.05	88.00	89.99	90.09	91.78	90.88	90.88

) is to obtain optimized input and output values required for relatively inefficient projects to become efficient. These values represent the amount of output or input resources required to achieve improvement. Such values are obtained using the DEA model, and they are referred to as Pareto-optimal solutions. However, the values should be used to identify an optimal direction for improvement and not to indicate actual goals to strive for. Whether managers can amend them according to their optimal target requires consideration of limited resources and the difficulty of improving the actual situation; resource allocation and improvement priorities can be readjusted.

4.2. Construction and Evaluation of ANN Model

Prediction models can be categorized into classification and regression. Classification involves assigning samples to predefined categories, where the labels of the target variables indicate the membership of the samples in these categories. Classification algorithms identify datasets of feature variables and build models based on these datasets to predict the class membership of new samples with unknown labels. The primary purpose of regression analysis is to predict a continuous numerical target variable. Unlike classification models, which allocate samples to specific categories, regression models focus on estimating the precise values of the target variable.

ANN can be used for both classification and regression tasks, depending on the nature of the problem and the network design. In classification tasks, ANN assigns data to different categories, typically utilizing Softmax or Sigmoid activation functions in the output layer to generate a probability distribution for each category, such as determining whether a construction project is classified as good or bad. In regression tasks, ANN predicts continuous numerical values, with the output layer commonly using a linear activation function to produce real numbers over any range, such as scoring a construction project.

The highest improvement ratio of inefficiency in the 12 airport projects was found to require a reduction of 54.76% in the frequency of defects. This demonstrates that defects are a significant factor in affecting project performance. In this study, 113 defects and inspection scores from renovation projects in the PCBMS are used as the features and labels, respectively. An ANN-based classification and regression model is constructed to evaluate the impact of defects on project quality. By using different layer structures and varying the number of neurons in each layer, and fine-tuning various hyperparameters, the learning capability of the model is improved to enhance the accuracy of predictions.

The ANN structure used in this study is composed of one input layer, three hidden layers, and one output layer. (1) The input layer: The data from renovation projects are converted into numerical format to serve as input features, facilitating easier network training. The input data includes both defects and inspection scores. (2) The hidden layers: This study utilizes three hidden layers, with the number of neurons in each layer being 200, 100, and 50, respectively. The more neurons and layers that are used in ANN, the more closely it can approximate complex continuous functions. ReLU is used as the activation function, as it has a wide range of output values and it has a faster gradient descent rate, thus it reduces training time. (3) The output layer: The number of neurons in the output layer depends on the nature of the task. For instance, binary classification and regression problems require only one neuron, whereas multi-class classification problems require multiple neurons. In this study, the classification network employs the Sigmoid activation function to classify projects into two categories: those with scores above 82 and those with scores below 82, indicating the quality of the construction projects. Additionally, the regression network uses a linear activation function to predict the score for each input project. The purpose of the activation function is to convert the output into binary values or probabilities.

In this study, the ANN model is trained using a batch size of 30 for gradient descent updates. An “epoch” refers to a single training iteration using the entire training dataset. The classification model requires multiple epochs to accurately predict the training data. In this study, the model is trained for a total of 200 epochs and achieved a training accuracy of 100% (as seen in Figure 5). The testing accuracy of the model on unknown data is 94.1%. The loss function used for training is binary crossentropy, and the model was trained for 200 epochs. The loss decreased rapidly in the initial training, but stabilized and reached convergence at around the 125th epoch. The final training loss was 0, as shown in Figure 6. The ANN model effectively captures the relationship between defects and project quality and can be used for decision making and analysis.

The most important consideration in model development is to evaluate its performance without bias. To achieve unbiased estimates, the dataset is divided into a training set (80%) and a testing set (20%). The training set is used for model development, while the testing set is used for performance evaluation. Training and testing are essential processes in executing supervised machine learning techniques. Figure 7 shows the confusion matrix results of the classification model for 17 airport renovation projects.

Table 7. Performance evaluation of ANN classification model.

Metrics	Accuracy	Precision	Recall	F1-Score
ANN Classification Model	0.941	0.667	1.0	0.8

shows that the ANN achieved the best predictive results regarding construction quality in the projects, with evaluation metrics of accuracy, precision, recall, and F1-score being 94.1%, 66.7%, 100%, and 80%, respectively.

The regression model employs four evaluation metrics: mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and R². The MSE is 0.11, indicating that the model’s prediction error is relatively small, demonstrating a certain level of predictive capability. The MAE is 0.22, which is also relatively low, suggesting that the average error in the model’s predictions is minor, reflecting good performance. The RMSE is 0.33, showing that the model’s prediction error is not high, and its units are consistent with the original data, making the evaluation results more meaningful. The R² value is 0.42, meaning that the model can explain approximately 42% of the variance. Although this is not very high, it still indicates that the model captures trends in the data to some extent (Figure 8).

4.3. Sensitivity Analysis

In neural network models, hyperparameters such as the number of layers, number of neurons, batch size, and dropout rate can significantly impact model performance. This study presents the results of a sensitivity analysis (see

Table 8. Results of sensitivity analysis for the ANN model.

Model	Accuracy	MSE	Number of Layer	Number of Neurons	Batch Size	Dropout
A	0.813	0.276	1	200	10
B	0.857	0.135	2	200 × 100	10
C	0.897	0.169	2	200 × 100	10	0.5
D	0.915	0.131	3	200 × 100 × 50	10	0.5
E	0.941	0.110	3	200 × 100 × 50	30	0.5

), which shows that as the number of layers increases, the model learns more complex features, thereby improving the accuracy and reducing the mean squared error (MSE). This explains why Model A (1 layer, accuracy 0.813, MSE 0.276) performs the worst, while Models D and E (both with 3 layers) achieve higher accuracies (0.915 and 0.941, respectively) and lower MSEs (0.131 and 0.110).

Furthermore, the number of neurons determines the model’s capacity, as a larger number of neurons allows the model to learn more feature information. The configurations of Models D and E (200 × 100 × 50) further enhance their learning capabilities, demonstrating optimal performance in terms of accuracy and MSE. The choice of batch size also affects the model’s training stability; Model E utilizes a larger batch size (30), which results in a more stable training process and, consequently, the best predictive capability.

Finally, the implementation of dropout techniques, particularly at a rate of 0.5, effectively mitigates overfitting, thereby improving the model’s performance on test data. This explains why Models C, D, and E outperform Models A and B, which do not employ dropout. In summary, the careful adjustment of these hyperparameters can significantly enhance the model’s predictive ability and generalization performance.

5. Conclusions

Facing the dual challenges of resource allocation efficiency and project quality requirements in current construction management, this study employs DEA and ANN methods to evaluate the performance of airport renovation projects in Taiwan. These methods effectively address the complexity of multiple inputs and outputs, providing a comprehensive data analysis framework. The main findings indicate that inefficient projects need to reduce defect frequencies by 54.76%, and both progress and budget implementation efficiency must be increased by 10.33% to achieve efficiency. Additionally, the classification accuracy of the ANN model reached 94.1%, and the mean squared error (MSE) for regression predictions was 0.11, demonstrating the effectiveness of these methods in predicting construction quality.

This study fills a gap in multi-attribute decision analysis within construction performance evaluation by providing a feasible framework that integrates the multi-attribute performance evaluation methods of DEA and ANN, thereby enhancing decision support capabilities in construction management. This framework not only offers a new perspective for academia but also provides a foundation for practical industry operations. Construction managers can use the performance evaluation indicators presented in this study to reallocate resources for inefficient projects and optimize the management of key variables. Specifically, project managers can identify priority areas for improvement based on the DEA analysis results and use the ANN model to predict the quality of future projects, enabling earlier issue detection during the construction process and reducing risks.

This study enables practitioners to effectively identify and address bottlenecks in construction projects, thereby improving overall project performance. By applying the methods outlined in this research, managers can make more informed decisions, ensuring that projects are completed on time, within budget, and with a high standard of construction quality. By integrating DEA and ANN methods, this study offers a comprehensive performance evaluation model for construction projects, significantly enhancing the decision-making capabilities of construction managers. It provides an important theoretical foundation and practical guidance for the industry.

A limitation of this study is that it considered only items available in the PCBMS data for inclusion in the DEA model for efficiency evaluation. Therefore, the results indicate relative efficiency among DMUs. Simultaneous comparisons under the CCR model overlook scale differences between DMUs and render unable adjustment to the most favorable evaluation condition to determine whether the DMUs have increasing returns to scale, decreasing returns to scale, or constant return to scale. For a more comprehensive performance evaluation, more data should be collected, and an appropriate combination of inputs and outputs should be selected. Although all DMUs in this study were homogenous airport renovation projects, some external factors may have affected their homogeneity. For example, the location, construction season, and construction period of the projects should be carefully assessed to avoid reducing comparison accuracy. In addition, costs, scale, time difference, and construction complexity may have caused slight variations in the evaluation results. Future studies may use other evaluation techniques, such as ratio analysis, regression analysis, and the AHP or analytic network process (ANP), to evaluate the performance of airport construction projects, compare methods, and explore the effects of the correlation between inputs and outputs on efficiency to implement effective construction project management.