A thorough literature review on risk management using various methodologies and strategies is presented in this section. Also, Table 1 provide the summary of the related work with key area focused.

In this study, Roy (2023), a risk matrix is used to analyse a risk assessment that is based on recognised hazards. The study’s findings address the difficulties and possible gains from using ML in the building sector. The research highlights the need for expertise in order to comprehend datasets that are special to a certain project. Examining issues with data consistency that impact data dependability, the research mainly focuses on unstructured text and image data. Despite the study’s acknowledgement of ML’s ability to digitalise and simplify construction procedures, it highlights obstacles, including data security [11] .

This paper (2022) places an emphasis on predicting software project failure early on as a means of risk assessment. Various methods of ML will be used. Machine learning is used in the development of the model. LR, NB, SVM, DT neural networks, and adaptive neuro-fuzzy inference systems were chosen as six methods to diversify the model. This work advances the area of software system development by creating models for software project risk assessment that are broadly applicable to any software project at any stage of the software development lifecycle [12] .

This study Elokby et al., (2021) enforced project risk management procedures and made the IT project successful in Egypt’s telecom and IT industries. There were four metrics used to evaluate the success of the IT projects: Project scope, project quality, project cost, and project time (schedule). Identification of risks, preparation for managing those risks, evaluation of those risks (both qualitative and quantitative), preparation for responding to those risks, actual execution of those responses, and monitoring of those results are all parts of risk management. In order to fulfil the goals of this study, a questionnaire was created as the primary means of gathering primary data [13] .

This study Owolabi et al., (2020) suggests a method for predicting completion risk using Big Data Analytics predictive modelling. There are linear regression, regression tree, RF, SVM, and DNNs were built and validated for a completion risk predictive model using the dataset of 4294 PPP project samples delivered across Europe between 1992 and 2015. The conclusion and result of the study prove that, with a less average test prediction error compared to other traditional regression tools, random forest could be a valuable tool in predicting delays in PPP projects. Other aspects included in this study are the questions related to the choice of model, its training and validation [14] .

This paper Mahdi et al., (2020) offers a review of current literature regarding the establishment of methodological innovations in the area of ML for software risk analysis. The analysis of this review has also highlighted some patterns in the methodologies of ML, size measures, and the findings that have shaped and advanced progress of ML in project management. Besides, this study provides a better understanding and a profound foundation for more research about software project risk assessment work. Additionally, it offers an additional method to minimise the likelihood of failure and improve the software development performance ratio, and it increases the likelihood that a software project will be anticipated and prepared to handle [15] .

The article Burkov et al., (2020) takes into account the responsibility of overseeing projects’ and programs’ hazards. Qualitative risk assessments are often used in practice to determine the degree of effect and the likelihood of harm. A three-point risk scale (low, medium, and high) is the most often used. The approach to managing risks in projects and programs, which relies on qualitative

evaluations and employs methods to avoid or reduce risks, have not evolved enough, in our opinion. This article provides an overview of risk aversion and risk reduction, several ways to accomplish these objectives, and a method for determining the qualitative features of potential dangers [16] .

As shown in Figure 1 , there are different phases of the project risk management and all the stages are clearly explained below.

A process starts with a large-scale data gathering campaign, with data on as broad a set of project history as possible. The collected data extends but is not limited to the following, project schedules, separate task schedules, and resources assigned to each task and project, and the results of these projects. This kind of collection of data is important since most of the aspects of project management are complex and may involve different factors that might affect a success of a project.

Data preprocessing meaning pertains to the processes that are taken through to clean data and make them fit for other uses. However, before using it in machine learning algorithms, a set of operations has to be performed in order to enhance its quality. Data normalisation, consistency checking, and managing missing values are essential parts of data preparation. Here are the main pre-processing techniques:

In cases where numerical inputs are used in feeding data, a feature standardisation process is conducted so that all inputs are normalised. This means that elements with large relative sizes cannot dominate the learning process which is very important for models that depend on the scale of the characteristics. To standardize a numerical attribute $aj$ Compute the standardized value $a_{ij}^{\text{*}}$ as follows (1):

$a_{ij}^{\text{*}}=\frac{a_{ij}-{\mu }_{aj}}{{\sigma }_{aj}}$ (1)

where ${\mu }_{aj}$ is a mean and ${\sigma }_{aj}$ is a standard deviation of attribute $aj$ across all projects.

The use of feature engineering with t-Distributed Stochastic Neighbour Embedding (t-SNE) was utilised to decrease a data dimentionality in order to retain its essential structure. This made it easy to determine features that are important to projects and whose absence would greatly affect project results, helping build better predictive models.

t-Distributed Stochastic Neighbor Embedding (t-SNE) was chosen over alternatives like Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for its ability to handle non-linear relationships between features while preserving the local structure of the data. This is crucial in project management data where task dependencies, resource allocation, and project outcomes often follow non-linear patterns.

As part of the experimental setting, the dataset was divided into two parts: The training set, which included 70% of the data used to train the models, and the testing set, which had 30% of the data.

Belonging to the family of ensemble learning methods, the GBC is an effective ML approach. By integrating boosting and gradient descent, it generates a trustworthy prediction model. One common approach to iterative optimisation is to use DT to create a set of fundamental forecasting models, which are then fed into GBC [17] . It is possible to modify the learner models and the loss function simultaneously. In order to minimize the loss function, gradient boosting is applied to certain data samples. This method uses gradient descent to reduce the loss to a minimum. Equation (2) shows that the technique minimise the loss function.

$\hat{F}\left(x_i\right)-\underset{f\left(x_i\right)}{\min }\underset{i=1}{\overset{n}{{\displaystyle \sum }}}ℒ\left(y_i,F\left(x_i\right)\right)$ (2)

The observed value is denoted as $y_i$, and the model formed by merging the weak learners is denoted as $\hat{F}\left(x_i\right)$ [18] .

A performance matrix comparing real observations with model predictions was used to assess an effectiveness of a chosen models. Some of the criteria included in the performance matrix were Recall, F1-score, precision, and accuracy. A following metrics were computed for various classes: A higher number of True Positives (TPs) indicates that positive cases were properly identified, while a lower number indicates that True Negatives (TNs) were correctly classified negative occurrences. those that are mistakenly classified as positive are known as False Positives (FPs), while those that are improperly classified as negative are known as False Negatives (FNs). The chosen performance metrics—accuracy, precision, recall, and F1-score—are all directly related to practical outcomes in project risk management.

The following formulae may be used to represent the assessment metrics:

Accuracy: An 85% accuracy indicates that the model correctly identifies project risks in the majority of cases, which provides project managers with reliable information on potential risks. It enhances decision-making and reduces the chances of having to address emerging issues temporally. The formula (3):

$\text{Accuracy}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}$ (3)

Precision: Precision of 82% suggests that when the model predicts a project risk, it is mostly correct. The following formula of precision (4):

$\text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}}$ (4)

Recall: Recall of 85% shows the model’s ability to identify actual project risks among all risks. In combination, they fine-tune the accurate identification of potential risks with low numbers of false positive findings. The formulation of recall is (5):

$\text{Recall}=\frac{\text{TP}}{\text{TP}+\text{FN}}$ (5)

F1-score: Thus, it shows the balance between the risk identification and avoiding false positive/negative values of 80% by the F1-score, which is the harmonic mean of precision and recall. This balance is crucial for project managers as they need accurate predictions that lead to effective intervention without overburdening resources. The F1-score formula is (6):

$\text{F1-Score}=\frac{2\ast \text{Precision}\ast \text{Recall}}{\text{Precision}+\text{Recall}}$ (6)

As a means of aiding decision-making and providing an objective measure of the models’ performance, this study used these four measures to assess the ML models.

The experimental findings reveal the great benefits of adopting the use of predictive analytics in project management. A significantly higher level of classification accuracy is shown by the Gradient Boosting Machine (GBM) model, with the accuracy estimated at 85%, precision at 82%, and recall at 85%, which proves the high efficiency of this model for further accurate prediction of project outcomes. The gradual increase in value of model accuracy as displayed above shows that the model can be tweaked in future hyperparameter tuning iterations with resultant better performance metrics. Additionally, predictive analytics achieves a resource utilisation efficiency of 85%, compared to 70% for traditional allocation methods, and a project cost reduction of 10%, double the 5% achieved by traditional approaches. The real-world data presenting the outcomes of the projects as plotted in the scatter plot of the predicted against actual results has shown a lot of congruity of the results of the predictive model. Overall, these findings are closely associated with the notion of how big data can change the nature of concurrent decision-making and improve the existing approaches to project management.

This section provides a comparative analysis between ML models for project risk management. The ML models are GBM, LR [12] , and SVM [12] that compare across performance parameters.

Figure 6 shows a comparison of model performance. The GBM achieved

the highest accuracy at 85%, with a precision82%, a recall 85%, and an F1-score 80%. On the other hand, the LR model reported fewer accuracy at 71%, but it had the best precision at 83% and F1-score at 87% to guarantee the method’s FP and FN. The SVM, provided a good performance with an accuracy 83%, precision84%, recall 77% and an F1 measure 83%. These findings imply that even though GBM is most accurate, LR provides better trade-off among precision and recall and is, therefore, more suitable in conditions where true positives are significant.