Data-driven methods in risk assessment of public-private partnership projects Open Access

Infrastructure plays a vital role in the development of societies, supporting both social and economic progress (Lu et al., 2022; Mortezaei Farizhendy et al., 2020). In fact, it acts as the essential foundation for sustainable progress (Akomea-Frimpong et al., 2023a,b). Many countries are facing challenges with ageing infrastructure, such as roads and bridges. Despite the growing demand for new infrastructure projects, governments worldwide face financial constraints that hinder their ability to fund all necessary upgrades and new initiatives (Li et al., 2022; Osei-Kyei and Chan, 2019). The necessity for significant investment in infrastructure is widely recognised, as evidenced by the McKinsey Global Institute's estimation that an annual expenditure of $3.3 trillion is required until 2030 to sustain anticipated worldwide growth rates (Burke and Lipshitz, 2024). Most countries are not spending enough money on infrastructure projects like roads, bridges, and utilities. This lack of investment creates a yearly $350 billion gap globally between the money needed for infrastructure and the money actually spent (Burke and Lipshitz, 2024). Because of the large financial gap, many governments are becoming more interested in partnering with private companies (public-private partnerships (PPPs)) to get additional funding sources (Jefferies, 2006). This approach can help cover the shortfall in infrastructure funding and facilitate the execution of public projects (Li et al., 2017). A PPP is described as a collaborative agreement between the public and private sectors, where each entity contributes its unique expertise and shares varying levels of risks and responsibilities with the goal of delivering efficient public infrastructure and services (Akintoye et al., 2003; Efficiency Unit, 2003).

Involving the private sector in public infrastructure procurement often leads to better project performance (Liu et al., 2015, 2017). The improved performance comes from the alignment of interests between public and private entities. This alignment, along with other incentives, encourages the public sector to involve the private sector in developing infrastructure. As a result, PPPs have become a favoured approach for delivering infrastructure, particularly for projects that require significant investment (Batra, 2023; Neto et al., 2020). Even though there is evidence from literature showing that PPPs are more effective than traditional procurement methods, where the public sector alone manages project delivery (Raisbeck et al., 2010; Xiong et al., 2017), there are issues with PPPs due to the complexity of financial agreements, long contract durations, investors' expectations for substantial profits, and the high-risk nature of PPPs (Yong Kim and Thuc, 2021). Furthermore, when risks are not properly allocated and the public sector shifts most project risks onto the private sector, it can negatively affect the goals of the project (Ahmadabadi and Heravi, 2019). PPPs have certain inherent traits, such as significant upfront capital investments, that make them prone to high risks of failure and can lead to major financial losses for both the public and private partners involved (Koc, 2023). For example, Abdullah and Khadaroo (2020) point out that lengthy contract durations, the participation of government authorities, and complicated relationships between public and private sectors are examples of traits of PPPs that can lead to unsuccessful completion.

If a PPP encounters risks such as financial instability, project delays, or contractual disputes, it can damage the reputation of private companies and leave public authorities facing unexpected troubles (Demirel et al., 2019; Zhang and Xiong, 2015). This significantly impacts both parties involved in the contracts (Wang et al., 2021). Being able to predict potential risks of PPP projects before they start would be very beneficial, as it allows for developing strategies to prevent potential risks. Consequently, there is a pressing need for an accurate and reliable model to predict and evaluate the various risks associated with PPP projects (Wu et al., 2020). Such a model would provide valuable insights, enabling decision-makers to identify potential issues early on, implement effective risk mitigation strategies, and ultimately enhance the overall success and sustainability of PPP initiatives.

This research provides a systematic review that evaluates previous studies on PPP risk assessment and highlights limitations, data sources, data-driven methods, and significant risk factors. By identifying gaps in the literature and suggesting improvements, this study offers insights for more accurate risk assessments in PPP projects. This study is novel in its scope, addressing a gap in the current literature by integrating PPPs, risk factors, and data-driven approaches across various fields, such as transport, waste management, and water systems. To the best of our knowledge, such an examination has not yet been fully explored in this context. The integration of artificial intelligence (AI) into the assessment of PPP contract failures significantly enhances decision-making and early-stage project planning (Liu et al., 2021). This approach demonstrates AI's effectiveness in forecasting the probability of contract failures, providing a quantitative analysis that supports more informed strategic decisions (Wang et al., 2021). It enables a thorough examination of the root causes behind potential failures, facilitating the development of robust risk mitigation strategies (Akomea-Frimpong et al., 2023a). Moreover, AI's predictive insights offer valuable improvements to the initial feasibility studies of PPP projects, enabling a more accurate evaluation of project feasibility. Consequently, decision-makers are able to make informed decisions and avoid high-risk projects, optimise resource allocation, and increase the success rates of future PPP hazards (Wang et al., 2021).

This research aims to answer the following research questions: (1) What are the significant risk factors affecting PPPs in data-driven studies? (2) What data sources have been utilised in previous data-driven studies to assess risks associated with PPPs? (3) What methods have been employed in previous studies on the implementation of data-driven approaches in PPP risk assessment? (4) What are the challenges of using data-driven approaches in the risk assessment of PPPs? The rest of this paper is structured as follows: Section 2 gives our methodology for the literature review. Section 3 examines the results provided by this study. Section 4 discusses the challenges and limitations of PPP risk assessment. Section 5 provides the discussion of this research study. Finally, Section 6 summarises the study's findings and implications.

The literature on risk assessment in PPPs using data-driven methods has been reviewed. For this purpose, Systematic Literature Review (SLR) steps are applied to the literature on PPP risk assessment. Web of Science (WoS) and Scopus are two of the largest databases that have been used to identify relevant papers. The steps of SLR are summarised in Figure 1. Step 1, using the research strings provided in Table 1, 288 manuscripts from WoS and 426 from Scopus have been obtained. In the second step, we narrow our scope to relevant subject areas. We have selected 11 fields in Scopus and 41 fields in WoS based on the relevance to the scope and these fields are related to these subject areas: (1) Engineering (2) Environmental and Earth Sciences (3) Computer Science and Information Technology (4) Business, Management, and Economics (5) Health and Medical Sciences (6) Agricultural and Biological Sciences (7) Social Sciences and Humanities (8) Multidisciplinary and General Sciences (9) Mathematics and Decision Sciences. As a result, 131 manuscripts from WoS and 211 from Scopus have been excluded. In the third step, only articles and reviews are selected. The primary reason for limiting the selection to articles and reviews is to ensure a focus on high-quality and peer-reviewed publications. This approach aligns with other SLRs, such as Akomea-Frimpong et al. (2023a, b), who similarly restrict their study to articles and journals. Rasheed et al. (2022) also exclude non-journal articles in their review. Additionally, many studies exclude grey literature, which includes some conference articles, as demonstrated by Salehi et al. (2021) and Ayub et al. (2020). Consequently, 16 manuscripts from WoS and 64 from Scopus have been excluded. In step four, non-English papers are excluded. As demonstrated in other SLRs, such as those by Rybnicek et al. (2020) and Borges et al. (2021), limiting the scope to English-written papers can ensure consistency and clarity in analysis. In the fifth step, an eligibility check is conducted, on the remaining 136 from WoS and 148 from Scopus based on these three criteria: (1) duplicate papers are removed from WoS to ensure a unique set of studies, (2) Scimago Journal Rank Q1 papers are selected to ensure the highest academic rigour and reliability, and lastly (3) the relevance of each paper to the research scope is confirmed by reviewing their titles and abstracts, and when necessary, the entire paper. Following these steps, 30 articles have been included for further analysis.

Figure 2 demonstrates the trend of publishing papers in PPP risk assessment using data-driven approaches. Since there is an increase in the trend, we can conclude that researchers are getting more interested in this area due to its importance. As shown, the largest number of published papers belongs to the period of 2022–2023. The growing adoption of PPPs for public infrastructure projects aligns with the financial constraints faced by governments, particularly in developing countries (Tan and Zhao, 2019). The subsequent section will define the significant risk factors associated with PPPs.

Based on the International Organization for Standardization definition in 2018, risk refers to how uncertainty can impact goals and objectives. In comparison to other types of projects, PPPs may carry more degrees of risk due to the involvement with many stakeholders, complicated arrangements, absence of experienced partners (Owolabi et al., 2020), and having specific regulations related to funding, documentation, and taxation (Wang et al., 2019).

There are numerous risks associated with PPPs. According to Rybnicek et al. (2020), the most significant risks can be classified into eight categories: contract, resources, objectives, structure, commitment, environment, communication, and trust. The factors mentioned most frequently are presented below.

1. Contract: Contracts are agreements between partners that bind them together and are a crucial part of every project. Contract-related issues can pose significant challenges in PPPs. There are three key risks with contracts: (1) negotiation, (2) incompleteness, and (3) design.

2. Resources: In PPPs, resources are valuable assets. These tools help partners work together, integrate their strengths, and tackle challenges. From the literature, finance, staff, and time problems are the three main risks related to resources.

3. Objectives: Objectives in PPPs encompass strategies, predictions, goals, or plans along with expectations for output quality in a project. Conflicting goals, strategic issues, and unclear objectives are key risks concerning project objectives.

4. Structure: Structural aspects deal with how PPPs are established and organised, and how partners collaborate effectively. Three main risks in this scope include allocation of responsibilities, decision-making, and coordination.

5. Commitment: Commitment is about how strongly people identify with the PPP and its objectives, how loyal they are to it, and if they are keen on putting in enough effort.

6. Environment: The risk factors that externally affect PPPs are environmental risks. These risks are difficult for PPP partners to control or can only be partly controlled by them. Several types of environmental risks include political risks, risks related to demand and revenue and risks associated with a competitive environment.

7. Communication: Communication risks involve sending the correct information to the right person at the correct time. There are three main communication risks: partner interaction, sharing information, and communication at the right time.

8. Trust: Trust is when one partner expects the other to behave properly, like not taking advantage of each other. There are three main trust risks that contain monitoring, trust building, and transparency.

Table 2 provides a summary of the most significant risk factors considered in the 30 studies on PPPs. Table 2 reveals that all eight significant risks were addressed, which indicates that this set of studies provides a representative sample of existing research on risk management in PPPs. This highlights the relevance and comprehensiveness of these studies in understanding the diverse risks associated with PPPs. Furthermore, Table 2 offers valuable insights for future research by underlining the significant risk factors that have been prioritised across different studies. As highlighted by Jeerangsuwan et al. (2014), the inherent diversity in the nature, scale, type, stakeholders, and location of projects makes it impractical and redundant to consider all risk factors during project assessments. Therefore, Table 2 serves as a practical reference for future researchers, guiding them in selecting a combination of significant risk factors relevant to their specific studies. In the following section, the data sources employed in these 30 papers will be elucidated and categorised.

Data from different sources has been gathered to investigate risks associated with PPPs. The most relevant data sources of the papers are presented in Figure 3. Primary data consists of the raw observations and measurements gathered directly by the researchers carrying out a study. It is collected with the specific goals of that study in mind, rather than relying on previously existing datasets. Primary data could encompass data from observations (for instance, those acquired via simulations), and data gathered through survey methods (like questionnaires). On the other hand, secondary data refers to data that is not directly obtained by the researchers. It could involve internal data amassed by government or private companies and shared with researchers for their study. Alternatively, it could involve data from external sources, like databases or scholarly papers (Sohail et al., 2023).

Primary data obtained from questionnaires has been used in previous studies (Ghorbany et al., 2022; Tariq and Zhang, 2021; Ye et al., 2018). Another type of primary data that has been employed by Erfani et al. (2021) is the data related to PPP projects. In their study, they collected data related to transportation PPPs and analysed it. The other type of primary data that has been applied is simulation. For example, Xiong et al. (2016), have utilised Monte Carlo simulation to assign probability distributions to the uncertain social-economic and project performance variables based on historical data.

Based on the papers collected in this research study, secondary data can be internal or external. For the internal data, there are two types, which include data from governmental or private organisations. Winata and Gultom (2023) use both government and private documents in their study. Regarding the external data, there are two groups that contain databases and articles.

For databases, there are different ones like the World Bank's Private Participation in Infrastructure database, World Governance Indicators, and World Development Indicators that have been used by some researchers. For instance, several studies have utilised databases in their research (Cheng et al., 2023; Koc, 2023; Tariq and Zhang, 2021; Wang and Tiong, 2022). Articles as a type of secondary data have been used by previous research studies (Ghorbany et al., 2022; Pu et al., 2021; Wang et al., 2022). In the forthcoming section, the data-driven methods employed in the 30 studies will be categorised.

Figure 4 depicts the methods used for risk assessment in PPPs by the collected papers. The methods fall into three main groups: Traditional statistical models, machine learning (ML) methods, and deep learning (DL) methods. Since the performance of a particular model significantly depends on the data it employs, the environmental factors, and the specific domain of the issue, it is challenging to make a thorough analytical comparison of current models. Consequently, in various situations, distinct models exhibit superior performance compared to other models. Therefore, selecting an appropriate method for risk assessment in PPPs requires a thorough understanding of the various parameters and scenarios involved.

In the context of PPP risk assessment, traditional statistical methods have been widely applied to analyse data related to project risks. Traditional risk assessment methods often rely on traditional statistical models. These methods exclusively depend on basic statistical assumptions, like the historical mean and k-sampling techniques. However, they lack the capability to incorporate external features related to risk assessment, limiting their ability to reduce uncertainty in the analysis. However, these models are widely used in real-world applications because of their conciseness and fairly precise results.

For example, Wang et al. (2019) employ Tobit regression to examine how the relationship between governance environment and risk allocation influences private investment. The results show that less risk assumed by private partners encourages more private investment in PPP projects. Moreover, the governance environment moderates the relationship between risk allocation and private investment. Specifically, better governance reduces the negative impact of risk assumed by private partners on private investment. Similarly, Zhang and Tariq (2020) conduct an extensive literature review to identify previously reported failure factors in water PPPs. They then do an in-depth analysis of 35 failed international water PPP projects using the event sequence mapping technique to map the sequence of events leading to failure. This allows them to identify failure mechanisms like chains of failure events triggered by an initial failure factor.

Furthermore, Xiong et al. (2016) develop a model to estimate fair compensation for early termination PPP projects. Linear regression and nonlinear regression are used to estimate performance variables and maximum likelihood analysis determines distributions for social-economic variables. The illustrative example shows the model can estimate a reasonable compensation amount and range for a hypothetical early terminated project. These examples highlight how traditional methods have been applied in real-world PPP projects to address risk-related challenges. However, these models often fail to capture the complex, dynamic risks inherent in PPP projects due to their reliance on basic assumptions and linear relationships.

AI methods, particularly ML techniques, have been utilised to improve the prediction and management of risks in PPP projects. ML methods have gained significant popularity in addressing a wide range of real-world problems (Ayodele, 2010). ML techniques can be effective in predicting risk factors in PPPs, such as the risk of failure in PPPs, inefficiency of PPPs, financial risks, and so on. Consequently, these models surpass complex mathematical models in performance.

Several popular ML methods employed for PPP risk assessment encompass Random Forest (RF) (Koc, 2023; Wang et al., 2021; Wang and Tiong, 2022), Regression Trees (Chou et al., 2016), support vector machine (SVM) (Chou et al., 2016; Wang and Tiong, 2022), example-dependent cost-sensitive (ECS) model (Wang and Tiong, 2022), Artificial Neural Network (Chou et al., 2016; Jin, 2011) and Neuro-Fuzzy model (Jin, 2011). For example, Koc (2023) develops an AI model to predict whether PPP projects will succeed or fail. The model demonstrates high accuracy and has the potential to enhance PPP project screening and resource allocation. Similarly, Wang et al. (2021) apply ML models to predict PPP contract failure across three sectors: transportation, water and sewer, and energy. These models, including Logistic Regression (LR), SVM, and RF, are compared using various data balancing techniques.

In another study, Wang and Tiong (2022) propose ECS ML models to forecast PPP contract failures. Their model involves a cost matrix based on misclassification costs and uses customised loss functions within gradient boosting and neural network algorithms. Chou et al. (2016) examine key factors influencing PPP disputes and apply various ML models, such as artificial neural networks (ANN), SVM, and Classification and Regression Trees, to predict the occurrence, type, resolution methods, and stages of these disputes. Jin (2011) proposes a neuro-fuzzy decision support system for efficient risk allocation in PPP infrastructure projects. This system, grounded in transaction cost economics and the resource-based view, uses neuro-fuzzy techniques to predict optimal risk allocation strategies.

Due to access to vast datasets and significant computational resources, DL methods have gained immense popularity over recent years. DL strategies have demonstrated remarkable effectiveness in addressing complex problems in the real world. One such method, Natural Language Processing (NLP), includes techniques like Word2Vec, which Erfani et al. (2021) applied to measure the similarity between risks identified in PPP and non-PPP projects in the US. This approach allows for better comparisons and insights into risk patterns across transportation projects. Overall, the application of AI models in risk-related studies within the context of PPPs has demonstrated their ability to deliver accurate risk predictions, which can enhance the likelihood of project success.

AI models tend to act like “black boxes”, where their inner workings are opaque (Lyngdoh et al., 2022). Thus, they typically lack the capability to offer insights regarding the extent of relationships among input data (Chehreh Chelgani and Makaremi, 2013). To better comprehend how these models arrive at their outputs, additional interpretation techniques are needed. XAI tools are particularly valuable in the context of PPP risk assessment because they provide insights into how various risk factors influence project outcomes. This clarity facilitates better decision-making for stakeholders involved in PPPs. One visualisation tool that can help is Shapley Additive Explanations (SHAP), which reveals how sensitive the model is to different inputs and shows the internal relationships between inputs and outputs. Using SHAP, the influences of input variables on model behaviour can be better understood (Lundberg et al., 2020). Moreover, as explained by Białek et al. (2022) when assessing the influence of features on PPP projects, the SHAP algorithm offers several distinct advantages in contrast to traditional methods for determining feature importance. SHAP feature analysis has been used in a few research studies (Ghorbany et al., 2022, 2023). In the subsequent section, the limitations encountered by the authors of the 30 papers will be discussed.

In examining the challenges related to data collection and quality within PPP risk assessment studies, several important issues highlight the need for thorough data collection and diverse datasets. For instance, Ghorbany et al. (2022, 2023) identify the lack of a comprehensive PPP project database that details performance conditions, complicating efforts to conduct in-depth evaluations. Similarly, Chen et al. (2016) encounter barriers with data limitations for empirical analysis, highlighting a common issue in PPP research, which is the lack of access to high-quality and extensive data sets. Jin (2011) faces limitations due to the relatively small number of PPP projects in the available dataset, while Han et al. (2019) highlight the challenges posed by insufficient data in risk research, stressing the need for more comprehensive databases. Thus, the reliability of findings also comes into question with small sample sizes, as demonstrated by Han et al. (2019) and Jin (2011). These examples collectively underscore the ongoing need for improved data collection practices and the development of more extensive datasets to advance the field and provide stakeholders with richer insights. So, future studies need to make efforts to collect widespread and sufficient datasets, as larger datasets tend to yield more accurate and reliable models.

This section highlights the existing limitations of some traditional regression methods in PPP risk assessment and the potential of AI techniques for improving model accuracy and flexibility. In the reviewed paper of this study, both traditional and AI methods are utilised by the researchers. For instance, studies like those by Cheng et al. (2023) and Wang et al. (2019) have used Tobit regression, which is limited by its assumption of linear relationships that do not capture the complexities inherent in PPP risks. Specifically, this limitation can lead to oversimplified risk analyses that fail to account for nonlinear interactions and dependencies among variables. In contrast, AI models like neural networks can capture complex, nonlinear interactions more effectively, as noted by De Marco and Mangano (2013). Moreover, Jin and Doloi (2008) have indicated the shortcomings of linear regression analysis in fully capturing the intricacies of PPP risk factors. AI methods, with their ability to learn directly from data, offer a promising alternative by reducing the need for manual feature engineering and potentially offering deeper insights, as seen in research by Kamugisha et al. (2019) and Koc (2023). Advanced AI techniques, not constrained by such assumptions, could provide a more robust approach for predicting PPP contract failures and refining failure mechanisms, as explored by Wang et al. (2021). Therefore, the transition to advanced AI in PPP risk assessment research can offer a more dynamic and flexible approach that better accommodates project complexities and improves predictive accuracy. Employing state-of-the-art AI models in the scope of PPP risk assessment research has the potential to provide accurate and more informed results and address current limitations.

In PPP risk assessment research, subjectivity and bias can influence the reliability and objectivity of study findings. For example, Chen et al. (2016) highlight how personal judgement in categorising studies could introduce biases, indicating that different researchers might organise the same studies differently. Similarly, Ghorbany et al. (2022, 2023) have to depend on expert opinions due to the lack of comprehensive data, which raises questions about potential biases influencing the results. In addition, Pu et al. (2021) note that PPP studies are predominantly found in certain fields, suggesting that the perspectives of authors might influence research categorisation and potentially introduce bias. Furthermore, Tariq and Zhang (2021) manually code data from case studies, a process susceptible to personal bias, suggesting that using AI for text mining could offer a more unbiased approach to identifying key themes. These examples underline the importance of minimising subjective influences in research to enhance the validity and reliability of findings in PPP risk assessment studies. Thus, implementing automated data analysis techniques could enhance the validity and reliability of findings by reducing human biases and increasing objectivity. Efforts should be made to avoid collecting data in ways that could introduce bias, which helps maintain the credibility of the results.

The lack of transparency in PPP risk assessment studies using AI techniques hinders the interpretation of their findings. For instance, Erfani et al. (2021) use DL to analyse risk similarities in transportation projects, but do not use XAI techniques like SHAP to make the AI results clearer. Similarly, Rudžianskaite-Kvaraciejiene et al. (2015) employ RF to evaluate PPP project efficiency from multiple perspectives, but do not incorporate XAI to easily interpret the ML results. Moreover, Chou et al. (2016) explore PPP disputes using various ML models. However, the study could have benefited from XAI techniques to clarify how these models arrived at their predictions. While these studies provide valuable insights, they could benefit from incorporating XAI methods. Integrating techniques like SHAP into these studies could enhance result clarity, making it easier for stakeholders to understand and apply AI-driven insights in PPP risk management. Therefore, it is suggested that future studies utilise SHAP or other techniques like that to interpret the results of AI and improve its transparency.

The generalisability of findings in PPP risk assessment research often faces limitations due to the regional focus of datasets. Studies like De Marco and Mangano (2013) and Erfani et al. (2021) are confined to the UK and the US, respectively, suggesting a need to include data from a wider range of countries for broader applicability. Similar constraints are seen in the work of Liu et al. (2024) focused on and Owolabi et al. (2020), which targets UK construction firms. Both studies indicate that exploring data from additional regions could enhance the relevance of their conclusions globally. Similarly, research by Zhang et al. (2023) underscores how country-specific factors can significantly influence PPPs. This geographical limitation suggests that expanding research to include diverse international settings could greatly enhance the relevance and applicability of PPP risk assessment findings on a global scale. Thus, focusing on multiple countries and areas rather than only one can result in developing models that are suitable for a broader area. This way, the developed models can help more project managers from different projects to make more informed decisions and manage risks in a better way.

The categorisation of data sources, methods, and limitations in the context of PPP risk-related studies has been done for the first time in this review, which provides a valuable contribution. This work can guide future studies by offering directions on significant risks in PPPs, applicable methodologies, data sources, and potential limitations researchers might encounter.

Effective risk management, which includes identifying potential risks and uncertainties, assessing their likelihood and impact, and then developing appropriate responses to address the risks, is essential for infrastructure projects to succeed. Effective risk management highly relies on proper risk identification since it forms the foundation for subsequent steps like evaluating, responding, and allocating risks (Jung and Han, 2017). Thorough and detailed identification of risks helps reduce the negative effects of unexpected circumstances in the future (Siraj and Fayek, 2019). Due to the importance of this topic, an increase in the number of research papers in this area is observed. In addition, this trend is likely to continue in the coming years because governments are struggling to allocate enough money for the increasing need for infrastructure due to the ongoing pandemic (Jallow et al., 2021).

LR can serve both as a traditional statistical model and as a component in ML models. However, in ML-based LR, there are distinct advantages compared to statistical LR, particularly concerning its ability to select the penalty type (L1, L2, or ElasticNet). Additionally, ML-based LR benefits from features that help capture complex relationships, including nonlinear ones (Pedregosa et al., 2011).

NLP techniques and DL methods have shown remarkable potential in various domains, but their application to risk assessment in PPP (PPP) projects remains largely undiscovered. The literature review highlights that only the study by Erfani et al. (2021) has explored this area, employing Word2Vec, a popular NLP technique, to measure the similarity of risk items across transportation projects. However, this is only the beginning, and there is a significant need for further research in leveraging the power of NLP and DL for understanding and assessing risks in PPP projects.

DL models, such as recurrent neural networks and transformer-based architectures like Bidirectional Encoder Representations from Transformers, have demonstrated great performance in NLP tasks, including text classification, sentiment analysis, and information extraction. These models can automatically learn complex patterns and representations from large amounts of unstructured textual data, such as risk documents, without the need for extensive manual feature engineering. By utilising the capabilities of DL, researchers can potentially develop more accurate and robust risk assessment systems that can effectively analyse and interpret the intricate risk factors present in PPP project documentation.

Research has consistently shown that AI models outperform traditional approaches in terms of accuracy and performance. For instance, Khanmohammadi et al. (2023) find that AI modelling techniques significantly exceed the predictive accuracy of traditional methods such as LR. Similarly, Wei et al. (2019) demonstrate that AI models are both comprehensive and effective across various forecasting horizons and applications, particularly excelling in delivering precise short-term energy consumption forecasts at the regional level. Furthermore, Niu et al. (2019) highlight that three AI methods, ANN, extreme learning machine, and SVM, demonstrate superior simulation performance compared to multiple linear regression. Additionally, Suzuki et al. (2019) conclude that ML techniques possess the potential to develop risk prediction models with robust predictive capabilities and effective discrimination of risk impact. This evidence underscores the advantages of employing AI models in enhancing risk assessment practices, particularly within the context of PPPs.

Generally, documents related to risks, like risk registers, are filled with lots of text and many different risk items, and the project team members explain the risks using their own terminology and words. As a result, manual comparison demands significant time and resources, but utilising advancements in NLP offers an automated tool for efficiently handling this comparison. NLP allows computers to understand and process human language data. NLP techniques convert text information into numerical formats that can then be utilised for data-driven modelling and analysis (Di Giuda et al., 2020). NLP has recently been applied more in various construction fields. For instance, it simplifies the review of contract terms and reduces the effort and mistakes from people having to manually read contracts. Furthermore, one important topic for researchers who are using NLP is automatically extracting contract clauses (Lee et al., 2020). Researchers and scholars employ NLP for different purposes such as document classifications, automatically analysing building codes and regulations (Zhang and El-Gohary, 2016), and exploring stakeholder views on social media platforms (Xue et al., 2020). Furthermore, NLP can serve as an automated tool to examine cases involving construction defects in court by analysing the frequency of keywords (Jallan et al., 2019). Knowledge extraction from data to make informed decisions is becoming increasingly popular in the construction industry today (Baker et al., 2020).

However, one of the main challenges associated with DL models is their “black box” nature, which makes it difficult to interpret and understand the reasoning behind their predictions or outputs. This lack of interpretability can be a significant drawback, especially in critical decision-making scenarios like risk assessment in PPP projects, where transparency and explainability are crucial for building trust and facilitating informed decision-making. To address this challenge, researchers have developed techniques for explainable AI (XAI), which aim to provide insights into the decision-making process of complex models like deep neural networks. One such technique is SHAP, which can help understand the relative importance of different input features and how they contribute to the model's output. By employing SHAP or similar XAI methods, researchers can gain valuable insights into the key risk factors identified by the DL models and how they influence the risk assessment process. While the study by Erfani et al. (2021) employs DL methods for analysing risk documents in PPPs, they do not utilise XAI techniques like SHAP to interpret and explain the results of their models. Incorporating SHAP or similar XAI methods could have provided valuable insights into the key risk factors identified by their models and how they influence the risk assessment process.

In conclusion, the application of NLP and DL techniques to risk assessment in PPP projects presents a promising avenue for future research. By leveraging the power of these advanced AI techniques, researchers can develop more accurate and robust risk assessment systems that can effectively analyse and interpret the intricate risk factors present in PPP project documentation. However, it is crucial to address the interpretability challenge by adopting XAI methods like SHAP, which can provide transparency and explainability, leading to better decision-making and successful project outcomes in the context of PPPs.

The implications of this review for theory are for future researchers working in the area of PPP risk assessment. This review has provided insights into the data-driven methods employed by other researchers and highlighted their limitations, thus providing a clearer understanding of the current landscape. Furthermore, it suggests the adoption of state-of-the-art AI methods in the risk management of PPP projects, which can encourage more effective approaches in future research. Additionally, this research serves as a valuable guide for selecting significant risk factors and types of data sources, helping to inform the design of upcoming research.

The practical implications of this review highlight the growing role of data-driven methods, particularly AI, in enhancing risk assessment for PPP projects. The application of AI techniques can improve decision-making by offering more accurate and timely insights, which facilitate better risk identification and mitigation. Traditional risk models may struggle to capture the complexity of PPP risks, whereas AI, combined with XAI, provides transparency and actionable insights. This combination of AI and XAI can lead to more efficient, accurate, and informed risk management, which can ultimately improve project outcomes and enhance risk management for both public and private stakeholders.

This SLR has examined the state of data-driven methods for risk assessment in PPP projects. Several key findings and implications emerged:

1. There is an increasing trend of published research applying data-driven and AI methods to assess risks in PPPs over recent years, which aligns with the growing adoption of PPPs globally for delivering public infrastructure projects. Therefore, this study can provide a foundation for future studies to identify emerging topics.

2. Researchers have applied a range of traditional statistical models, ML techniques like random forests and neural networks, as well as some initial DL methods to PPP risk assessment problems. However, the use of advanced DL and NLP techniques remains limited so far, which indicates potential for further exploration in the area of PPP risk assessment.

3. Key challenges identified include data quality and availability issues (covering both primary and secondary sources), subjectivity and bias in data collection, limitations in model specifications that fail to fully capture complexity, and difficulties generalising findings beyond specific contexts. Thus, future researchers can address these challenges, derived from the literature on PPP risk assessment using data-driven approaches, to achieve more general and possibly enhanced results.

4. NLP and advanced DL architectures represent a promising direction to better leverage large volumes of unstructured text data (contracts, reports, case studies) for assessing PPP risks. However, interpreting these complex models requires incorporating XAI methods. Integrating XAI methods like SHAP can help improve the transparency and interpretability of the “black box” ML and DL models applied to PPP risk assessment. Only a few of the reviewed studies employed XAI techniques. This demonstrates that relying solely on AI methods may be insufficient, and integrating XAI alongside them can provide both researchers and practitioners with more transparent and interpretable insights.

In summary, while data-driven approaches offer advantages for PPP risk assessment, there are still gaps in employing the latest AI techniques, especially DL and NLP combined with XAI interpretation methods. Overcoming data limitations and developing transparent, robust AI models tuned for the PPP context is crucial for more reliable risk assessment to improve public infrastructure delivery through PPPs. This review highlights opportunities for further research integrating cutting-edge AI techniques into PPP risk analysis.

One of the limitations of this study is the focus on PPP studies. By concentrating exclusively on this area, the review may miss valuable insights that could be gained from examining a broader scope, such as research related to Alliances or other integrated project delivery models. Future studies could benefit from expanding the search string to include other relevant procurement methods. Furthermore, this review specifically categorised data-driven methods in PPP risk-related studies. Future research could expand its scope by categorising all methods utilised in PPP risk assessment, rather than restricting the focus to data-driven approaches. Moreover, Q1 journal papers are selected for this study to ensure a focus on high-quality, peer-reviewed publications. However, future studies can investigate other studies that were not the focus of this study to improve generalisability.