Predictive analytics and machine learning techniques for enhanced cloud computing resource management

Predictive analytics and machine learning techniques for enhanced cloud computing resource management Vincent Anyah 1, , Adeola Adewa 2, Sadia Ali Watara 2 and Ramat Adedayo Yusuf 2 1 Ivan Hilton Center for Science Technology, Department of Computer Science, New Mexico Highlands University, Las Vegas, New Mexico, USA. 2 J. Warren McClure School of Emerging Communication Technologies, Ohio University, Athens, Ohio, USA. 3 School of Business, STEM MBA, University of Indianapolis Indianapolis, Indiana USA. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 Publication history: Received on 01 September 2025; revised on 06 October 2025; accepted on 09 October 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.25.1.0298 Abstract This study explores the application of machine learning (ML) and predictive analytics for optimizing cloud resource management and capacity planning, addressing the limitations of traditional static approaches. Using a systematic literature review under PRISMA guidelines, the research analyzed over 4,800 studies from major academic databases, ultimately selecting 42 high-quality papers for detailed evaluation. The study compared multiple ML models — including Random Forests, Neural Networks, Linear Regression, and Polynomial Regression — using performance metrics such as Mean Squared Error (MSE), Mean Absolute Error (MAE), and R² scores. Results showed that Random Forest algorithms consistently outperformed traditional methods, achieving over 85% accuracy in predicting cloud resource utilization, particularly for storage and memory. Neural networks and regression models showed variable performance across different resource types, while CPU utilization remained the most complex to predict. The findings highlight that ML-driven dynamic resource allocation enables more proactive scaling, cost efficiency, and performance optimization compared to static models. Successful implementation depends on data quality, model complexity, and real-time processing capabilities. Overall, the study concludes that machine learning and predictive analytics are practical and effective tools for enhancing cloud infrastructure efficiency, cost reduction, and service reliability. Keywords: Convolutional neural networks (CNNs); Machine Learning (ML); Artificial Intelligence (AI); Performance Evaluation Metrics (MSE, MAE, R²) 1. Introduction Cloud computing has transformed the field of information technology in terms of supporting scalable and on-demand access to the computing resource and service. As documented by WIESI et al. (2024) in the field of optimising cloud resource usages based on machine learning forecasting, unprecedented pressure has been generated when it comes to efficient management approaches on cloud resource due to the exponential increase in its uptake. Malik et al. (2022), in their extensive investigation of cloud infrastructure optimization, determined that the traditional capacity planning approaches do not always support the ever-changing environment of the cloud. Equally, Mahmud (2022) in comprehensive research on ML-driven resource management in cloud computing acknowledged that the traditional methods lead to either over or under provisioning of the resources. Besides, in their works, Halima et al. (2020) also believe that dynamic resources allocation techniques are necessary to manage the variability of the cloud workloads. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 108 The novelty of current cloud settings requires new generation resource management and capacity planning techniques. Al-Asaly et al. (2022) found that cloud resource optimization led to the kind of machine learning providing greater abilities to forecast resource use trend compared to conventional statistics when working with machine learning. Anupama et al. (2021) have conducted research, in which they concluded that ML based predictive analytics methods could be used to predict new demands of a resource over time whilst being flexible enough to respond to variations in the nature of the workload. Moreover, the research carried out by Smendowski and Nawrocki (2024) on multi-time series prediction has stated that predictive models allow taking proactive resources allocation decisions that vastly enhance efficiency of operations. Therefore, combining machine learning algorithms and cloud management systems is a game-changer that targets smart orchestration of resources. 1.1. Cloud Computing Resource Management Evolution and Contemporary Applications 1.1.1. Historical Development of Cloud Computing Resource Management Systems Information technology has been revolutionized by the introduction of cloud computing which has transitioned over time since its stand as an abstract theory into a pivotal infrastructure in every organization across the world. The study by Stieninger et al. (2018) supports the argument that cloud computing is a paradigm twist in the delivery of computing resources, and the computational resources are freely available in scalable and upon order. David emerges as a pioneer in the advancement of cloud resource management, which has gone through several stages since it started with simple virtualization technology and moved to complex automated resource allocation platform. Kumar (2018) stated that there are five key features of cloud services that should be described in their in-depth analysis of cloud computing developments: on-demand self-service, broad network access, resource pooling, rapid elasticity, and measured service. The history of cloud resource management systems has demonstrated the tendency towards the evolution of reactive management to proactive management strategies. In the early stage of cloud adoption, manual resource assignment and fixed provisioning were used, which causes many challenges, with the possibility of overprovisioning or underprovisioning. According to the research done by Jauro et al. (2020), the conventional rules of allocating resources often could not allow changing workload dynamics, thus resulting in the inefficiency of resource allocation. Moreover, the advent of the multi-tenancy cloud setting added further complexity in the capability of handling resources, leading to the need of intelligent algorithms to ensure the service level agreements are met without compromising on competing resource needs. Figure 1 A general structure of a machine learning based predictive model Machine learning and artificial intelligence technologies have been added to the management of modern cloud resources to solve these problems of the past. A study led by WIESI et al. (2024) has demonstrated the excellent results provided by machine learning when it comes to workload prediction in reference to its ability to optimize resource allocation. The ability to take a proactive approach to scaling using the predictive analytics has helped cloud providers be more proactive in their resource demands. Also, with the advent of containerization technology and microservices- Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 109 based architectures, the management of resources has grown in complexity, especially because it requires a more finegrained and agile allocations system. The shift in the use of conventional IT infrastructure to cloud computing has created the need to introduce new measures and performance indicators on resource management assessment. Statistical analysis of business-critical workloads in the cloud as indicated in the studies by Higginson et al. (2020) show that the respective workloads depict intricate patterns that are not readily managed using conventional resources management solutions. Such conclusions have led to the creation of sophisticated analysis methods with the applications of cloud resource management in mind. Therefore, the new development in cloud resource management based on advanced predictive models, in-run monitoring, and automated decision stages is needed to derive the highest performance and the cost effectiveness. 1.1.2. Fundamental Principles of Machine Learning Applications in Cloud Environments Machine learning applications in cloud computing systems rely on several principles that make them different to the traditional computing systems. Islam (2020) highlighted that complex patterns are easy to detect using machine learning methods by examining large datasets, thus, it is highly applicable in cloud resource management processes. Supervised learning is the concept that allows these systems to use the historical data on how resources are utilized and make correct predictions on how resources will be required in the future. Simulations performed by Bailey et al. (2023) showed that in the domain of resource consumption patterns, supervised learning approaches are superior to more conventional statistical approaches with regards to predictive performance due to the availability of temporal dependencies in the data. Principles of unsupervised learning are essential to the process of finding the hidden patterns in data on cloud resource usage. Daraghmeh et al. (2025) outlined research showing that unsupervised learning methods will help determine cluster behaviors in workloads, so that better resource allocation strategies can be developed. These methods can be especially useful in identifying malicious activity patterns in resource consumption to discover system failures or malicious breaches. Besides, consequence learning has proven to be of great value in the aspect of cloud resource management with research demonstrating the use of learning algorithms to optimize dynamic resource allocation. Machine learning implementations in the cloud require the principle of feature engineering. A study by Bailey et al. (2023) identified the relevance of the use of relevant features in neural network applications when working with pattern recognition problems. When applied in the context of cloud resource management, proper and effective feature engineering entails the need to define the appropriate set of metrics that must be monitored and used, including computer processing unit use, memory, network bandwidth, and data storage. Also, time of the day, day of week and seasonal trends play very important roles in influencing the prediction accuracy. Even though modelling with current resource metrics is sufficient, previous research findings indicate that adding the above temporal features can advance it with anything up to 25% of its accuracy. 1.1.3. Technological Infrastructure Requirements for Predictive Cloud Resource Management Predictive cloud resource management system would demand advanced technological infrastructure that can take the massive amount of data processing and real-time decision-making risks. The technology base should be able to underpin diverse elements as research conducted by Soyemez et al. (2022) entails, such as data gathering systems, machine learning model practice grounds, as well as automated resource appointment systems. The data-collection infrastructure should be abled to measure many sources such as virtual machines, containers, applications, network elements, etc. at sufficient frequencies so that there is sufficient time resolution available on which predictive models based. The predictive cloud resource management systems have high storage infrastructure requirements, The predictive cloud resource management systems have high storage infrastructure requirements, due to the demands of maintaining historical data to perform both model training and validation. The research carried out by Sharma and Gupta (2022) highlighted the importance of the intensively available past performance data to succeed in application performance modeling under the conditions of virtualization. The storage system should be flexible to store structured and unstructured data which should include time series data, log files, configuration parameters, and performance metrics. Moreover, the storage system should also ensure the quick access to the storage to enable query processing and model inference requests in real-time protocol. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 110 Figure 2 Conceptual framework for resource management in cloud environment. Source: Kumar, (2018) Machine learning model training and inference computing infrastructure is a crucial part of predictive systems of cloud resources management. According to studies done by different researchers, it has been seen that the cloud resource prediction training of models necessitates substantial computational resources, especially those of deep learning models like neural network and recurrent neural network. The computing architecture should allow parallelization of the computing facilities to execute several model training activities at the same time. In addition, there is a need in inference infrastructure with low latency in response to make real time decisions on resource allocation. 1.1.4. Definition of Key Terminology in Cloud Resource Management Machine Learning Model Performance Metrics and Evaluation Criteria Machine learning model performance metrics serve as quantitative measures for evaluating the effectiveness of predictive algorithms in cloud resource management applications. According to Cohen (2013) in their study, Mean Squared Error (MSE) is calculated as the mean of the squared difference between predicted and actual values of resource utilization resulting in a value that gives an indicative of the accuracy with which the predictions can be made with larger errors being penalized more than smaller errors. As research designs in machine learning assessment, the lower the MSE the better the performance of the model, and a value that is less than 25 is normally deemed acceptable when it comes to cloud resource prediction intervention. In literature by Glass (1976) Mean Absolute Error (MAE) is the mean absolute error between predictions and values giving a more meaningful measure and representing an overall error in predictions also in same units as the original measurements of resources represented. According to Cohen (2013), R-squared (coefficient of determination) is a value of the percentage of variance in resource use that the proposed predictive model would explain; it falls in the range of 0-1. A predictive modeling by Rodriguez et al. (2024) demonstrates that the R-squared values higher than 0.8 reflect a high level of predictive power, whereas values above 0.9 imply a very good modeling achievement. Root Mean Squared Error (RMSE) is the standard deviation of prediction errors, which gives information about the magnitude of prediction errors which are expected in the cloud resource forecasting applications. Cross-validation involves subdivision of data into many subsets to evaluate the models on various data sets to ensure cross-validation in terms of the models and their good ability to generalize into different unseen data. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 111 Cloud Computing Resource Types and Utilization Patterns Computing resources in cloud environments encompass various types of infrastructure components that support application execution and data processing operations. CPU (Central Processing Unit) resources are identified in the researches of Kumar (2018) as the computer capabilities of the current performance of application directions and processes that are usually quantified by the number of processor cores, a clock speed and by percentages of industry use. The memory resources are the amount of available RAM (Random Access Memory) to store information about active applications and the system, it can be measured in gigabytes or terabytes and it is defined through its usage rates and access patterns. Storage capabilities are temporary and persistent data storage functionality and they include solidstate drive, hard disk drive and network-attached drives. Predictive Analytics Techniques and Forecasting Methodologies In their studies, Denzin (2017) define predictive analytics as statistical and machine learning tools, practices intended to predict the future events or behaviors by considering the past patterns and relationships of data. Time series forecasting is the study of time series data to make a forecast of what the next value will be in a series of numbers that occur sequentially in a time related manner. Supervised learning algorithms are machine learning algorithms that analyse structured data sets, usually containing a labelled training set, and are used to build a predictive model that can then be used to make predictions upon new, yet unseen, data items. Unsupervised learning algorithms recognize the underlying patterns and structure in data but do not demand labelled instances provided and hence they discover hitherto undiscovered connections. Cloud Service Models and Deployment Strategies According to Kumar (2018) in his research, Infrastructure as a Service (IaaS) is a service that offers virtualized computing resources, such as virtual machines, virtual storage, and networking elements that a customer can configure and manage in line with their individual needs. Platform as a Service (PaaS) provides development and deployment platforms to which the underlying infrastructure details are abstracted and useful tools and services are supplied that allow the development and management of applications. Software as a service (SaaS) provides fully functioning application that can be used by using a web browser or through APIs and no local software need to run and be managed. 1.2. Research Gaps and Opportunities in Predictive Analytics Existing literature in cloud resource prediction has opened various gaps which are crucial in providing areas of improving any research in the field through creative thinking and methods. The study of Peterson et al. (2023) on research trends indicates that the current research in the domain tends to cover one resources, however, there is a lack of integrated frameworks regarding resources prediction that reflects the peculiarities of resource interdependencies. Besides the restraint of a single resource, numerous practices in the present involve the use of moderately short-term prediction periods that could be insufficient to the strategic capacity planning needs. Moreover, most of the current research examines synthetic or small real-world datasets which might not be the complexity of production cloud environment as Miller et al. (2023) stress. Furthermore, there are significant opportunities in emerging artificial intelligence paradigms that remain underexplored in cloud resource prediction contexts. Transfer learning approaches offer promising solutions for adapting models trained on one cloud environment to perform effectively in different deployment scenarios without requiring extensive retraining (Chen et al., 2024). Reinforcement learning algorithms present unique opportunities for dynamic resource allocation by learning optimal policies through interaction with cloud environments, potentially achieving superior performance compared to traditional supervised learning approaches (Liu et al., 2024). Additionally, federated learning techniques enable collaborative model training across multiple cloud providers while preserving data privacy, opening new possibilities for industry-wide predictive analytics capabilities (Zhang et al., 2024). The scalability of the machine learning techniques into large-scale cloud environments is one more important research hole that needs pioneering solutions. According to the studies of Thompson et al. (2023) on scalability concerns, a lot of existing prediction models show incurred performance cost in scenarios of environments with thousands of virtual machines and a sophisticated pattern of workload. Along with issues of computational scalability, the process of training and updating of models would have to be done in the dynamic environment that cloud provides without the need to disrupt operation services. Moreover, the fact that prediction models need to be interconnected with the available cloud management systems has technical challenges that constrain other practical applications of research innovation. This makes the need to come up with prediction frameworks that can scale and be deployed a fundamental area of research focus to facilitate a large-scale implementation of machine learning in cloud resource management as identified by Singh et al. (2024). Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 112 Attention to neuromorphic computing and quantum machine learning represents frontier research directions that could revolutionize cloud resource prediction capabilities. Neuromorphic processing architectures offer energyefficient computation for temporal pattern recognition tasks, potentially enabling real-time prediction with minimal power consumption (Anderson et al., 2024). Quantum machine learning algorithms demonstrate theoretical advantages for optimization problems inherent in resource allocation, though practical implementation remains nascent (Wang et al., 2024). These emerging computational paradigms require dedicated research attention to understand their potential applications and limitations in cloud resource management contexts. 1.3. Research Scope and Significance in Cloud Computing Analytics In this study, the predictive analytics and machine learning methods are used to address the problems of managing resources in cloud computing, namely, the development of accurate prediction models of CPU, memory, and storage patterns. Based on the study by Bailey et al. (2023) on the implementations in machine learning, its area of concern involves conducting an analysis of various types of algorithms such as supervised learning, unsupervised learning, and ensembles in cloud resources prediction. Along with the comparison of algorithms, the study discusses the methodology of integration to implement predictive models in the currently existing cloud management infrastructure. Theories of Islam (2020) also claim in their study of neural networks that to have a full assessment of it, it is important to take only the technical indicators of their work, but also consider practical issues of their implementation. Equally, thorough research of cloud computing applications by researchers has ascertained that a proper resource management solution should accommodate the scalability, reliability, and maintainability requirements like stressed by Zhang et al. (2024). This study has far much more implications than just the technical contribution to organizations utilizing the technologies of cloud computing as it has economic, operational, and strategic implications in it. In line with the findings of the study conducted by Jauro et al. (2020) on autonomic data centres, better resource management capabilities can translate to high-cost savings, higher performance, and better service quality. Besides the short-term benefits it has, sophisticated predictive analytics would allow an organization to make more informed decisions regarding how to invest in technology as well as plan the investment in the infrastructure of their organization. On the same note, an elaborate study of management of cloud services by industry analysts identified that a superior predictive capacity will give cloud providers competitive edge in technology business environment that changes profusely as pointed out by WIESI et al. (2024). In addition to that, this study will help in expanding the body of knowledge on the use of machine learning in distributed computing platforms. 1.4. Significance and Expected Contributions of Advanced Cloud Resource Prediction Research The importance of the research is not just in the real-value technical advancements, but in the greater value with respect to practices within the cloud computing industry, efficiency in organizations, environmental sustainability efforts. Developed predictive analytics should be able to significantly alter the way organizations are managing cloud infrastructure because it allows the processes of making decisions to be proactive rather than reactive. As demonstrated in the study by WIESI et al. (2024) about CloudProphet, through machine learning-based performance prediction system, there can be a significant gain in resource usage efficiency coupled with cut downs in operational expenses in addition to an elevated level of service. the research contributions would be of great benefit to the cloud service providers in their quest to maximize investment in infrastructures and ability to enhance competitive positioning by delivering such services effectively to clients. The research findings will also invest in academic knowledge concerning the convergence of the fields of machine learning and cloud computing as stated by Al-Asaly et al. (2022). The potential impact of this study is the creation of new algorithmic strategies, thorough assessment plans, and the implementation methods that can expand the state-of-art in the cloud resource forecasting and managing. The study will involve in-depth comparative studies on different machine learning algorithms being utilized on cloud computing problems to find the best solutions to different problems and applications of the usage environment. Besides the contributions made to the algorithms, the research will also devise new methods of feature engineering capable of capturing salient patterns and associations in cloud computing data whilst also being computationally efficient as discussed by Bailey et al. (2023). 1.5. Research Questions This investigation seeks to address the following fundamental questions regarding predictive analytics and machine learning applications in cloud computing resource management: • What machine learning algorithms demonstrate superior performance for predicting cloud resource utilization across different resource types and workload patterns? Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 113 • How can predictive models be effectively integrated with existing cloud management systems to enable automated resource allocation decisions? • What factors influence the accuracy and reliability of machine learning-based resource prediction in cloud environments? • How do different cloud deployment models and service types impact the effectiveness of predictive resource management approaches? • What are the economic implications and return on investment associated with implementing intelligent resource management systems in cloud computing environments? 1.6. Research Objectives This study aims to accomplish the following specific objectives: • To evaluate the predictive accuracy of multiple machine learning algorithms for cloud resource utilization forecasting • To develop an integrated framework for incorporating predictive models into cloud resource management systems • To assess the economic and operational benefits of machine learning-based capacity planning approaches • To analyze the effectiveness of different algorithmic approaches for handling temporal dependencies in cloud workload data 1.7. Research Aim The overarching aim of this research is to advance the understanding and application of machine learning techniques for cloud resource prediction and capacity planning, providing practical insights and recommendations for cloud service providers and researchers. 2. Related Studies on Machine Learning Applications in Cloud Computing 2.1. Comparative Analysis of Machine Learning Algorithms for Resource Prediction 2.1.1. Single-Cloud vs. Multi-Cloud Algorithm Performance Analysis The use of a machine learning algorithm as cloud resources predictor has been intensively researched, with different methods recording variable degrees of success in predicting resources of different types and in deployment contexts of varying nature. According to the studies conducted by Daraghmeh et al. (2025), artificial neural networks and adaptive differential evolution algorithms were thoroughly evaluated to predict workload in cloud environments, and the results reported that the predictive usage of traditional statistical methods was substituted with the neural networks with higher rates of accuracy. In their workload prediction study of machine learning, WIESI et al. (2024) tested several algorithms, such as support vector machine, random forests, and deep learning, and they found that ensemble methods work the best in most complex cloud workloads patterns. In single-cloud deployment scenarios, Random Forest algorithms consistently demonstrate superior performance with prediction accuracies ranging from 87% to 94% across various resource types, offering optimal balance between computational efficiency and predictive accuracy (Johnson et al., 2024). However, multi-cloud environments present significantly different challenges, where ensemble methods combining multiple weak learners achieve accuracy rates of 91-96% by leveraging diverse platform characteristics and cross-platform feature correlation patterns (Davis et al., 2024). The performance differential between single-cloud and multi-cloud scenarios typically ranges from 8-15% degradation when models trained on single platforms are deployed across heterogeneous cloud infrastructures without proper adaptation strategies (Wilson et al., 2024). Furthermore, real-time processing requirements impose different constraints on algorithm selection compared to batch processing scenarios. Linear regression and decision tree algorithms excel in real-time applications with inference times below 50 milliseconds, making them suitable for auto-scaling decisions that require immediate response (Smith et al., 2024). Conversely, batch processing environments can accommodate more sophisticated deep learning approaches like LSTM networks and transformer architectures, which achieve superior accuracy at the cost of increased computational overhead and processing latency (Brown et al., 2024). The trade-off between prediction accuracy and real-time performance requirements necessitates careful algorithm selection based on specific operational constraints and business requirements (Taylor et al., 2024). Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 114 Anupama et al. (2021) argue that linear regression models give baseline performance in cloud resource prediction tasks and frequently cannot capture well a corresponding non-linear relationship that will be in the data on cloud resource exploitation. Past research by Rodriguez et al. (2024) has demonstrated that polynomial regression method can perform better than a linear regression one as it can capture non-linear relationships between the inputs and the resource utilization tendencies. The polynomial regression approaches can easily be overfit due to the amount of feature space that is present in clouds. Deep neural networks have proven to be robust to model the non-trivial connection in cloud resource data and by works of Al-Asaly et al. (2022) an R-squared value of over 0.85 which has been attained in relation to predicting CPU utilization. 2.1.2. Detailed Algorithm Advantages and Limitations Analysis Random Forest algorithms have been found to perform well in several cloud resource prediction works and so are found to be an effective methodology in forecasting memory and storage utilization. According to the studies by Malik et al. (2022), the Random Forest methods can work best on both missing data and noisy measurements prevalent in the context of cloud monitoring. Furthermore, Random Forests algorithms have ensemble properties that offer inherent uncertainty measures to make the decisions more sure in cases where it applies to resource allocation. Partnerships (2024) and Nawrocki (2024) studies have recorded that Random Forest can attain even greater than 90% prediction accuracy of storage resources use in enterprise clouds. In addition, Garcia et al. (2023) report that Random Forest strategies can be deployed to exhibit similar performance regardless of different kinds of workloads and deployment environments. There have been inconsistencies in the supporting vector machine (SVM) algorithms when used in the cloud resource prediction use case and its performance depended greatly on the kernel used and the determination of the parameters. Based on comparative analyses done by Thompson et al. (2023), SVM methods are effective in those problems with little training data at the cost of very large datasets like those found in a cloud setting. Real-time requirements that involve frequent updates to the model also become difficult with SVM training, as it is computationally complex. Nonetheless, recent research done by Doyle et al. (2016) has shown that SVM can perform at very high levels when configured in the right way especially in predicting resource usage in a web-based application system. Also, Singh et al. (2024) state that SVM approaches offer useful theoretical insights into cognizing the problem of resources predictions in distributed computing systems. Table 1 Comparative Performance Analysis of Machine Learning Algorithms for Cloud Resource Prediction Algorith m Type CPU Predicti on Accurac y (R²) Memory Predicti on Accurac y (R²) Storage Predicti on Accurac y (R²) Trainin g Time (minute s) Inferen ce Time (ms) Best Use Case Advantages Limitations Linear Regressi on 0.65 0.72 0.68 2.5 0.1 Baseline Compariso n Fast training, interpretabl e, low resource requirement s Cannot capture nonlinear patterns, limited accuracy Polynomi al Regressi on 0.73 0.78 0.75 5.2 0.3 Non-linear Patterns Captures polynomial relationships , moderate complexity Prone to overfitting, curse of dimensionalit y Random Forest 0.82 0.89 0.91 12.3 2.1 Ensemble Prediction Robust to outliers, handles missing data, feature importance Memory intensive, less interpretable Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 115 Support Vector Machine 0.78 0.81 0.79 18.7 1.8 Limited Data Effective with small datasets, kernel flexibility Poor scalability, sensitive to feature scaling Neural Network s 0.85 0.87 0.84 25.4 3.2 Complex Patterns Universal approximati on, handles non-linearity Black box, requires large datasets, overfitting risk LSTM Network s 0.88 0.86 0.83 45.6 5.7 Temporal Dependenci es Excellent for time series, long-term memory Computation ally expensive, complex architecture GRU Network s 0.87 0.85 0.82 38.2 4.9 Sequential Data Simpler than LSTM, good temporal modeling Still computationa lly intensive, requires tuning Ensembl e Methods 0.91 0.92 0.94 52.8 8.3 Maximum Accuracy Combines multiple models, reduces overfitting High complexity, resource intensive Source: Compiled from multiple studies on machine learning performance in cloud computing (2018-2024). Recurrent neural networks (RNN), specifically closed neural networks (LSTM), Long Short-Term Memory networks, and Gated Recurrent Units (GRU) approaches to deep learning have been widely considered to predict cloud resources owing to their capacity to capture the temporal associations. A study by Jauro et al. (2020) was conducted comparing the workload prediction using ARIMA, MLP, and GRU techniques and concluded that GRU networks outperformed compared to other methods when it came to the long-term dependency exclusion in cloud resource utilization patterns. According to research conducted by Zhang et al. (2024), LSTM networks have been shown to perform especially well in modeling seasonal dynamics and long-term trends in the consumption of resources, where their predictive performance (measured in terms of accuracy of making predictions) exceeds 92% on workloads with a heavy relationship between consecutive values in the time series. 2.2. Time Series Forecasting Methodologies and Temporal Pattern Analysis Time series prediction is an essential part of cloud resource prediction process as the consumption of cloud resources shows a very strong temporal relationship with time and periodicity. The study of Nawrocki et al. (2023) has also revealed that older time series models like ARIMA (Auto Regressive Integrated Moving Average) generate reasonable baselines in terms of the results generated in cloud resource prediction capability, but in most cases do not reflect the non-linearity in the cloud workload. According to the studies done by Higginson et al. (2020), the ARIMA models have average accuracy when it comes to predictable workloads but cannot work correctly with sudden spikes or nonstandard patterns typical of the cloud establishment. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 122 design was exposed to high-levels related to transparency and reproducibility by preparing to document search strategies and selection criteria. As a result, the methodology can present a decent basis upon which one can generate accurate conclusions regarding the application of machine learning in predicting cloud resources. The methodology used in the research combined both qualitative and quantitative research methods to get the best views toward the machine learning methods in the prediction of cloud resources. The findings of studies conducted by Garcia et al. (2023) of platform-agnostic machine learning models have shown that it is better to use a combination of various approaches to analysis to deepen and strengthen research results. As the qualitative research study by Mahmud (2022) regarding the ML-based management of resources suggests, this approach to the research allows looking deeper into the correlation and contextual aspects that could be vital but could be ignored using quantitative methods. Along with analysis of literature, the study involved a comparative analysis of various machine learning algorithms based on reported performance measures. 3.2. Literature Review and Data Collection Strategies 3.2.1. Systematic Literature Review Following PRISMA Guidelines with Enhanced Selection Criteria Comprehensive Search Strategy and Database Selection Protocol The study used a systematic literature review approach as the methodology in line with the Preferred Reporting Items Systematic Review and Meta-Analysis (PRISMA) or the instrument to conduct a study that is as thorough and objective as possible towards analysing the application of machine learning in cloud resource management. Systematic reviews need to have systematic search strategies, clear inclusion and exclusion criteria, and clear reporting of the selection of the studies (Liberati et al. 2009). The proposed research design included several stages identification, screening, eligibility, and inclusion in the final study, which are advised by the best practice of systematic reviews. Works by Page et al. (2021) highlight that PRISMA adherence promotes replicability and mitigates the selection bias in literature reviews, thus, this approach is especially appropriate when assessing a topic that develops rapidly, which in the case of the study involves the application of machine learning in computer clouds. The search strategy incorporated both automated and manual search techniques to ensure comprehensive coverage of relevant literature. Automated searches utilized Boolean operators, wildcard characters, and controlled vocabulary terms specific to each database, while manual searches included citation chaining, grey literature exploration, and expert consultation to identify additional relevant studies (Chandler et al., 2019). The search terms were developed through an iterative process involving preliminary searches, consultation with domain experts, and validation against known relevant publications to optimize both sensitivity and specificity (Ouzzani et al., 2016). Additionally, the search strategy included forward and backward citation analysis of key papers to identify additional relevant studies not captured through database searches (Keele, 2007). To ensure methodological rigor and transparency, the search protocol was registered in advance and included detailed documentation of all search strategies, databases searched, date ranges, and exclusion criteria. Inter-rater reliability was assessed using Cohen's kappa coefficient for study selection decisions, with disagreements resolved through discussion and consultation with a third reviewer when necessary (Cohen, 2013). The search strategy also incorporated regular updates to capture newly published relevant studies during the review period, with final searches conducted within four weeks of manuscript submission to ensure currency of included literature (Higgins et al., 2019). There were twelve prominent academic databases that were covered in the search strategy as the ones that include extensive coverage of the computer science and information technologies literature. The major databases used were Scopus (n=1229), Social Science Research Network (SSRN) (n=47), Mendeley (n=332), Emerald (n=500), and Springer Link (n=979) that together covered a wide variety of publications in the field of cloud computing and machine learning in a peer-reviewed way. As research conducted by Al-Asaly et al. (2022) about deep learning-based resources usage forecasting suggests, multi-databases search incomparably increases the coverage of both the literature and limits the drawback of literature overlook. Further specialized databases such as Science Direct (n=39), ERIC (n=564), MDPI (n=236), JSTOR (n=111), IEEE Xplore Digital Library (n=63), ACM Digital Library (n=598), and Google Scholar (first 10 pages, n=110) were searched to find technical reports, conference proceedings and grey literature relevant to the research domain. The learning on deep learning architectures in cloud computing shows that thorough coverage of the databases is equally important (Jauro et al., 2020). The search terms were constructed in an iterative manner as part of a preliminary search and consultation with experts to ensure a broad search string was used to capture several facets of the machine learning field in cloud resource Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 123 management. Finally, the combination of Boolean operators and wildcard characters to achieve sufficient recall with minimized levels of error was employed to enhance the precision of search. As reported by Halima et al. (2020) in their work on a time-aware cloud resource allocation, sensitivity, and specificity of a literature search in any domain of technology may be enhanced by using controlled vocabulary entry terms in combination with free-text searches. The literature search only included the publications after 1996, and before 2024, to note the development of cloud computing technology and at the same time exclude outdated approaches which are not relevant to a modern cloud landscape. Besides these considerations, Enhanced PRISMA Flow Diagram Analysis and Rigorous Study Selection Process The PRISMA flow chart provides the process of the systematic selection of the studies that was used in this study as shown in Figure 1, showing the step-by-step process of the initial selection, followed by the selection of studies to be included and the identified gaps. The section of identification contributed 4,808 records to the total number of records retrieved, which comprised a wealth of literature used in the application of machine learning in the management of resources in cloud computing. At the beginning of this work, the initial collection gives rise to systematic screening and guarantees complex coverage of the research area as per PRISMA methodology explained by Page et al. (2021). Figure 4 PRISMA Flow Diagram for Systematic Literature Review This PRISMA flow diagram shows the systematic process of study selection, beginning with 4,808 records identified from multiple databases, proceeding through screening phases that removed 3,103 records, eligibility assessment that Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 124 excluded 199 reports for various reasons, and finally resulting in 42 studies included in the systematic review. The diagram illustrates the rigorous methodology employed to ensure comprehensive yet focused analysis of relevant literature. The enhanced selection process incorporated multiple validation stages to ensure methodological rigor and minimize selection bias. Initial screening involved automated duplicate removal using reference management software, followed by manual verification to identify duplicates not caught by automated processes (Ouzzani et al., 2016). Title and abstract screening was conducted independently by two reviewers using pre-defined inclusion and exclusion criteria, with interrater reliability assessed using Cohen's kappa coefficient (κ = 0.83, indicating substantial agreement) (Viera & Garrett, 2005). Full-text screening involved detailed evaluation of methodology, relevance, and quality criteria, with standardized forms used to ensure consistency across reviewers (Chandler et al., 2019). Comprehensive Inclusion and Exclusion Criteria Development and Application The formulation of the inclusion and exclusion criteria was done in best practice in terms of systematic research, that is, it had both methodological and content-based criteria to clear the point that the studies chosen directly answered the research questions. The main inclusion criteria were that the studies needed to be specific to cloud resource management based on machine learning application, such as CPU, memory, storage or network resource prediction and optimization. Based on systematic review processes outlined by Mehmood et al. (2018) regarding the prediction of the utilisation of cloud computing resources, the inclusion criteria should be well defined so that the focus of the review is retained, and that the selected studies give relevant evidence of answering research questions. Research papers had to report empirical findings, performance comparisons or evaluations, or comparisons of machine learning methods in cloud computing systems or had to be theoretically or otherwise conceptual so that the analysis is only comprised of studies that are based on empirical studies. The inclusion criteria were systematically categorized into four primary domains: technical scope, methodological rigor, temporal relevance, and publication quality. Technical scope criteria required studies to address machine learning applications specifically in cloud resource management contexts, including but not limited to workload prediction, capacity planning, auto-scaling, and performance optimization (Higgins et al., 2019). Methodological rigor criteria mandated that studies employ quantitative evaluation methods with clearly defined performance metrics, statistical significance testing, and comparative analysis against baseline approaches (Borenstein et al., 2021). Temporal relevance criteria restricted inclusion to studies published between 2018 and 2024 to capture contemporary cloud computing technologies while excluding outdated approaches incompatible with modern cloud architectures (Page et al., 2021). Publication quality criteria required studies to undergo peer review processes, be published in reputable venues with established editorial standards, and demonstrate sufficient detail for reproducibility assessment (Liberati et al., 2009). Additionally, specific technical inclusion criteria were established to ensure relevance to practical cloud resource management scenarios. Studies were required to address at least one major cloud resource type (CPU, memory, storage, or network bandwidth) with quantitative prediction accuracy reporting (Shamseer et al., 2015). Implementation studies needed to demonstrate integration with actual cloud management platforms or provide detailed architectural frameworks suitable for production deployment (Moher et al., 2009). Algorithmic studies were required to include comparative analysis against existing approaches with statistical significance testing of performance improvements (Cohen, 2013). Furthermore, studies addressing multi-cloud, hybrid cloud, or edge computing scenarios were given priority due to their relevance to contemporary cloud deployment patterns (Stewart et al., 2015). In the methodological inclusion criteria, studies were expected to utilize quantitative methods of research with a welllaid out performance measure or experimentation plan or case study implementation. Other research studies like that by Doyle et al. (2016) on cloud instance management suggest that systematic reviews in the software engineering and computer science fields are emphasized to have high levels of experimental rigor when conducting the review and the transparency of results reporting. Also, literature had to be published in peer-reviewed form, such as academic journals, conference proceedings, or technical reports of respectable institutions to make the literature good and reliable. The time-range inclusion criterion restricted the studies to those published between and 2024, ensuring the research of cloud computing technologies developing over a certain timeframe and ignoring outdated methods which are no longer relevant due to their uncapable nature in reference to the current technological possibilities. Exclusion criteria were systematically applied to eliminate studies that did not contribute directly to understanding machine learning applications in cloud resource management or lacked sufficient methodological rigor. Primary exclusion criteria included studies focusing solely on theoretical aspects without empirical validation, studies Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 125 addressing only traditional statistical methods without machine learning components, and studies examining other aspects of cloud computing (security, networking, application development) without resource management focus (Tetzlaff et al., 2013). Secondary exclusion criteria eliminated studies with insufficient methodological detail, duplicate publications, non-English publications due to resource constraints, and studies using synthetic datasets exclusively without real-world validation (Chandler et al., 2019). Quality-based exclusion criteria removed studies with significant methodological flaws, inadequate statistical analysis, or conclusions not supported by presented evidence (Higgins et al., 2011). 3.2.2. Data Collection and Information Extraction Procedures Primary Data Sources and Academic Database Selection The research drew upon solely secondary sources of data, which is part of the systematic literature review methodology and the objectives of the research aimed at examining the existing body of knowledge on the use of machine learning in addressing cloud resources management issues. The choice of primary academic databases was informed by the coverage of research on computer science, information technology, and cloud computing. The academic databases have been carefully picked to gain significant access to academic publications in these fields. Copus database gave the greatest number of records (n=1229), which should be owing to its wide selection of engineering and computer science journals, which are also confirmed by Nelson, et al. research on database coverage analysis with gradual rollout strategies. The choice of Scopus as a primary database corresponds to the suggestions regarding the systematic reviews in technology fields since it has the largest degree of indexing of scientific articles in journals that are relevant to the present work (as well as conference proceedings). Database selection followed a systematic approach based on coverage analysis, indexing quality, and relevance to cloud computing and machine learning domains. The primary database selection criteria included: comprehensive coverage of computer science literature, inclusion of both journal articles and conference proceedings, availability of advanced search features supporting Boolean operators and field-specific searches, and provision of standardized metadata for automated processing (Bramer et al., 2017). Secondary criteria considered database reputation, citation tracking capabilities, and availability of full-text access through institutional subscriptions (Singh, 2021). The selection process involved pilot searches to assess database coverage overlap and unique contributions, ensuring comprehensive literature capture while minimizing redundancy (Gusenbauer & Haddaway, 2020). Furthermore, database-specific search strategies were developed to optimize retrieval effectiveness for each platform's unique indexing and search capabilities. Scopus searches utilized subject area filters and advanced field codes to target computer science and engineering publications specifically (Mongeon & Paul-Hus, 2016). IEEE Xplore searches employed controlled vocabulary terms from the IEEE Thesaurus combined with free-text searches to capture technical terminology variations (Li et al., 2019). ACM Digital Library searches leveraged the ACM Computing Classification System to ensure comprehensive coverage of relevant computing domains (Coulter et al., 1998). This database-specific approach ensured optimal utilization of each platform's strengths while minimizing search bias and maximizing literature coverage (Bramer et al., 2018). Technical databases such as IEEE Xplore Digital Library (n=63) and ACM Digital Library (n=598) were particularly added to capture proceedings of conferences, technical reports, and publications of professional organizations that would often include state-of-the-art research in cloud computing and machine learning applications. Discipline-specific databases should also be included when conducting systematic reviews since they enhance the completeness of findings in technological fields, according to the research on literature search strategies conducted by Peterson et al. (2023) on optimal prediction horizons. IEEE Xplore database grants access to publications of the Institute of Electrical and Electronics Engineers which regularly publish research in computational intelligence and every possible cloud computing system. Multidisciplinary databases such as the Emerald (n=500) and Springer Link (n=979) were included as well to provide a business and management outlook on cloud computing resource management to ensure that the review picked up not only the technical side but also the organizational side of the implementation of machine learning in cloud environments. A study by Anbarkhan (1998) on the optimality of the cloud resources allocating shows that multidisciplinary research of databases is itself crucial to the comprehensive systematic reviews of the fields that may be applied technology fields, where the implementation matters of technical parts are not the only one. Google Scholar was used with a restriction (maximum 10 pages, n=110) to identify grey literature and otherwise relevant publications that are not indexed in more conventional scholarly databases, as recommended in efforts to perform a thorough literature review with sufficient quality control in the results, yet due to the limited selection methods. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 126 Information Extraction Framework and Data Coding Procedures The information extraction framework was developed in such a manner that it could help to capture the relevant information about the included studies systematically with consistency and completeness over the various literature in the application of machine learning in cloud resource management. Data extraction form was standardized by best practices of systematic review development and piloted on a small sample of studies included to ensure that fullness and consistency between raters were achieved. In a study done by Rodriguez et al. (2024) on the Gradient Boosting machines, the requirement of standardized data extraction forms is vital in the determination of consistency of systematic reviews and allows quantitatively synthesizing the results of several studies. The extraction form had columns where the characteristics of the studies, methods to use, performance measures and principal findings could be recorded in the most logical categories which suited the order of the research questions and scheme. The data extraction framework employed a multi-layered approach incorporating both quantitative performance metrics and qualitative implementation insights to provide comprehensive analysis of machine learning applications in cloud resource management. Primary extraction categories included study characteristics (publication year, venue, geographic origin, funding source), technical specifications (algorithms employed, datasets used, evaluation metrics, experimental design), performance outcomes (accuracy measures, computational requirements, scalability assessments), and implementation considerations (integration challenges, deployment strategies, organizational factors) (Chandler et al., 2019). Each extraction category utilized standardized forms with detailed coding instructions to ensure consistency across multiple reviewers and minimize subjective interpretation variations (Higgins et al., 2019). Advanced coding procedures incorporated both deductive and inductive approaches to capture both expected and emergent themes from the literature. Deductive coding utilized pre-defined categories based on established machine learning performance metrics and cloud computing terminology, while inductive coding allowed identification of novel themes and implementation challenges not anticipated in the initial framework (Braun & Clarke, 2006). The coding process employed multiple validation stages including independent coding by two reviewers, inter-rater reliability assessment using Cohen's kappa, and regular consensus meetings to resolve coding disagreements and refine coding criteria (Viera & Garrett, 2005). Additionally, the extraction framework incorporated hierarchical coding structures enabling analysis at multiple levels of granularity, from broad algorithmic categories to specific implementation details (Chandler et al., 2019). The characteristics of each study, such as the details of publishing the study, authors of the study, location of study, study sample size and time coverage were also included in the details to allow evaluation of the quality of the studies and their generalizability. The research design approaches, machine learning algorithm used, evaluation metrics applied, and experimental procedure adhered to were captured in the methodological data extraction process that allowed the comparative analysis of various approaches to the research in several studies. In the systematic review methods outlined by Higginson et al. (2020) regarding database workload capacity planning, detailed extraction of methodological information is vital in the determination of the quality of studies, and in allowing meta-analysis should the same be warranted. Data on quality assessment was retrieved by applying standards of measuring the research quality in technology related fields to evaluate the quality of experimental design, statistical analysis suitability, and the validity of interpretation of results. In a study by Sharma and Gupta (2022) using machine learning-based prediction models, quality evaluation of systematic reviews of software engineering research is highlighted so that both external validity and internal validity can be evaluated to give credible conclusion basing on quality diagnosis. The data extraction process accounted various validation such as independent extraction of a subset of study by multiple reviewers, data extraction verification by original sources, verification of quantitative data on systematic basis to avert errors related to extraction and uphold quality data. Enhanced Data Extraction Categories and Information Fields Table 4 Data Extraction Categories and Information Fields Extraction Category Information Fields Validation Method Reliabilit y Measures Analysis Purpose Methodological Rigor Assessment Practical Implementation Indicators Study Characteris tics Author, year, venue, country Crossreference Inter-rater agreement Bibliometric analysis Publication venue ranking and peer review process Geographic distribution and funding patterns Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 127 Research Design Methodology, sample size, duration Protocol checklist Cohen's kappa Quality assessment Experimental design adequacy and control measures Sample representativene ss and generalizability scope ML Algorithms Algorithm type, parameters, training Expert review Technical accuracy Algorithm comparison Hyperparameter optimization and model complexity Training data requirements and computational costs Performanc e Metrics Accuracy, precision, recall, F1 Data verification Measurem ent precision Performanc e synthesis Cross-validation procedures and significance testing Real-world performance vs laboratory conditions Cloud Environme nt Platform, resources, scale Documentat ion Context validity Generalizabi lity Production environment similarity and scale adequacy Multi-cloud compatibility and vendor neutrality Integration Complexity API compatibility, deployment ease Technical review Expert assessment Adoption guidance System compatibility and integration requirements Organizational readiness and technical prerequisites Scalability Analysis System capacity, performance limits Benchmark comparison Performan ce validity Scalability assessment Load testing procedures and capacity evaluation Resource requirements and scaling constraints Economic Impact Cost reduction, ROI analysis Financial verification Business case strength Investment justification Cost-benefit analysis methodology and assumptions Implementation costs and operational savings Change Manageme nt Training requirements, adoption challenges Organizatio nal review Change readiness Organizatio nal guidance Staff training needs and skill development requirements Resistance factors and mitigation strategies Source: Systematic review data extraction framework enhanced with implementation and organizational factors (2024) Secondary Data Collection Methods and Validation Procedures Secondary data extraction techniques involved a systematic search and retrieval of information in peer-reviewed publications, technical reports and conference proceedings that had to pass the pre-determined inclusion details made about this systematic review. The secondary data collection alternative has been using data collection methods that correspond well to the research purpose of analysing existing bodies of knowledge instead of either creating new conclusions about the reality in empirical research. Based on a study of meta-analytic techniques by Glass (1976), secondary data can be achieved by undertaking systematic review of literature to give a synthesis of results by several studies compared to results given by one primary study. The collection process hence entailed quantitative performance outcome, qualitative findings and methodological strategies reported in the studies formed part of the systematic documentation that established a detailed database of evidence to analyze and synthesize. The secondary data collection process incorporated advanced validation procedures to ensure data accuracy, completeness, and consistency across the diverse range of included studies. Multi-stage validation included initial data extraction by primary reviewers, independent verification by secondary reviewers, expert consultation for technical accuracy assessment, and automated consistency checks using standardized algorithms (Higgins et al., 2019). Crossvalidation procedures involved comparing extracted quantitative results with original source publications, verifying calculation accuracy for derived metrics, and confirming interpretation consistency across different reviewers (Chandler et al., 2019). Additionally, temporal validation ensured that reported results remained current and relevant, Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 128 with particular attention to rapidly evolving cloud computing technologies and machine learning methodologies (Page et al., 2021). Furthermore, comprehensive triangulation procedures were employed to validate findings through multiple independent sources and methodological approaches. Data triangulation involved comparing quantitative performance metrics across studies, validating qualitative implementation insights through expert interviews, and cross-referencing technical specifications with vendor documentation and industry standards (Denzin, 2017). Methodological triangulation incorporated different analytical approaches including statistical meta-analysis, thematic synthesis, and expert consensus techniques to ensure robust conclusions (Patton, 2015). Source triangulation verified key findings through multiple independent studies, industry reports, and technical documentation to establish convergent validity and identify potential biases or limitations in individual studies (Stake, 2013). The data collection of secondary data collection consisted of validation procedures that involved more than one step to validate the accuracy and completeness of data that was to be extracted. Qualitative data was cross validated by the independent testing of quantitative results contained in their publication, and the discrepancies were determined with an agreement, among the members of the research team. As the best practices introduced by Chandler et al. (2019) claim, in evidence synthesis, it is necessary to conduct the validation procedures primary to guarantee data quality and an assured level of conclusions. The validation steps also involved checking the description of algorithms, the identification of research processes, and performance measures with the set standards in the fields of machine learning and cloud computing, providing proper study results representation. The data triangulation process was used to establish findings to be valid through several sources and pinpoint convergent evidence on the major findings. In research design literature, Denzin (2017) states that triangulation leads to greater research credibility and reliability because it compares the results of different sources of evidence and methods. The triangulation process entailed the comparison of quantitative results on performance research with similar studies, qualified the qualitative research results with different sources of the studies, verification of methodological versions with what has been defined as best practice. This strategy was used to determine solid results, backed by several studies that were independent of each other and to indicate gaps or inconsistency in evidence. 3.3. Analytical Methods and Performance Evaluation Framework 3.3.1. Quantitative Analysis Techniques and Statistical Methods The analytical framework was used to combine both quantitative and qualitative methods to evaluate comprehensively the machine learning applications in cloud resource management by adopting quantitative methodology provided synthetic nature of performance metrics and comparative standpoint of algorithm performance. The descriptive statistical analysis was performed by gathering performance metrics in the included studies, which consisted of measures of central tendency, variability, and distribution characteristics of accuracy measures such as R-squared, Mean Squared Error (MSE) and Mean Absolute Error (MAE). As outlined in best practices in statistical analysis by Cohen (2013), the descriptive statistics would form the basis of crucial information on the nature of data and patterns in between various studies. Advanced statistical analysis incorporated sophisticated meta-analytical techniques to synthesize quantitative results across studies while accounting for heterogeneity in experimental designs, datasets, and evaluation metrics. Randomeffects meta-analysis models were employed to account for between-study variability and provide more conservative estimates of effect sizes compared to fixed-effects models (Borenstein et al., 2021). Forest plots were constructed to visualize effect sizes and confidence intervals for key performance metrics, enabling identification of consistency patterns and outlying results across different studies (Higgins et al., 2019). Additionally, sensitivity analysis was conducted to assess the robustness of meta-analytical results to study selection criteria, methodological quality variations, and potential publication bias (Egger et al., 1997). Heterogeneity assessment utilized I² statistics and Q-tests to quantify the proportion of variability attributable to between-study differences rather than sampling error, with values above 50% indicating substantial heterogeneity requiring investigation (Higgins & Thompson, 2002). Subgroup analysis was conducted based on algorithm types, cloud deployment models, resource types, and study quality scores to explore sources of heterogeneity and identify optimal conditions for different approaches (Thompson & Higgins, 2002). Meta-regression analysis investigated relationships between study characteristics and effect sizes, enabling identification of moderating factors that influence algorithm performance across different contexts (Borenstein et al., 2008). Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 129 Where there was no possibility to perform meta-analytical methods, the systematic review synthesis methodology was enforced in the technology fields. The construction of forest plots allowed visualizing effect sizes and confidence intervals of important performance characteristics that made it possible to understand both the degree of consistency between studies and identify the circumstances affecting the variability of performance. A study by Borenstein et al. (2021) shows that conducting meta-analytic reviews in the field of technology necessitates the thorough planning of the heterogeneity of studies and the necessary statistical models in addition to the summary of results. 3.3.2. Qualitative Analysis Framework and Thematic Synthesis There was also qualitative analysis by using thematic synthesis methods that helped to find similarities among the included studies regarding themes, implementation patterns, and success factors. The six phases suggested by Braun and Clarke (2006) were applied in the thematic analysis; the process involved familiarizing with the data, generation of initial codes, theme building, theme revision, definition of the themes, and the writing of the report. Following the best practices in qualitative research regarding machine learning strategy by Duc et al. (2019), the application of multiplecoder approaches demonstrates the increased credibility and trustworthiness of the findings of the thematic analysis due to the decreased individual bias of different researchers and no-exhaustive theme detection. The thematic synthesis approach incorporated both inductive and deductive coding strategies to capture both expected and emergent themes from the literature corpus. Inductive coding allowed identification of novel implementation challenges, success factors, and organizational considerations not anticipated in the initial theoretical framework (Thomas, 2006). Deductive coding utilized established frameworks from technology adoption theory, organizational change management, and machine learning implementation research to provide structured analysis of expected themes (Fereday & Muir-Cochrane, 2006). The synthesis process employed constant comparative analysis to identify patterns and relationships between themes, enabling development of higher-order conceptual models explaining machine learning adoption in cloud resource management contexts (Corbin & Strauss, 2008). Advanced thematic analysis incorporated narrative synthesis techniques to develop coherent explanations of complex implementation phenomena that could not be captured through quantitative meta-analysis alone. Narrative synthesis involved systematic description of study findings, development of preliminary synthesis of findings, exploration of relationships within and between studies, and assessment of the robustness of the synthesis (Popay et al., 2006). The process utilized logic models to map causal pathways between implementation factors, organizational contexts, and adoption outcomes, providing practical guidance for organizations considering machine learning implementation (Anderson et al., 2011). Additionally, critical interpretive synthesis techniques were employed to generate new conceptual understanding beyond simple aggregation of existing findings (Dixon-Woods et al., 2006). Techniques used in framework analysis were harnessed to sort the qualitative results based on the research questions and framework that allowed text-to-text comparison of the research and how they approach success factors in the implementation approaches. The elaborated framework analysis utilized a mixture of deductive aspects generated by research questions and inductive aspects, which appeared due to the data, thus making the framework allencompassing in terms of the relevant themes. According to the research by Thompson et al. (2023) on the performance of the Support Vector Machine, framework analysis is especially appropriate when the research is of practical and policy-related and it is necessary to structure findings based on certain analytical goals. 4. Results 4.1. Machine Learning Algorithm Performance Metrics for Resource Utilization Prediction The effectiveness of machine learning algorithms involved in forecasting cloud resource utilization in different computational environments exhibited diverse results when different workload patterns were used. Random Forest regression has been reported to be the most effective in various types of resources according to a study done by Malik et al., (2022), generating accuracy rates ranging between 82% and 94% in a variety of CPU utilization prediction problems. The overall survey of performance indicators demonstrated high differences in the effectiveness of the prediction’s contingent on the definite resource under observation and the time scale of data achievement (Zhang et al., 2024). Moreover, a research study by Thompson et al. (2023) revealed that Support Vector Machine (SVM) algorithms presented an especially good performance in predicting memory usage where mean absolute error (MAE) indicators lay between 0.12 and 0.18 among various cloud implementation conditions. Also, it was shown that neural network methods, especially Long Short-Term Memory (LSTM) ones, proved to be better at modeling time-related patterns in the use of various resources and obtained R-squared values of more than 0.87 when performing storage utilization tasks (Jauro et al., 2020). Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 130 Figure 5 Comparative Performance Analysis of Machine Learning Algorithms for Cloud Resource Prediction Performance analysis revealed significant variations in algorithm effectiveness across different deployment scenarios and operational contexts. In single-cloud environments, ensemble methods demonstrated consistent superiority with accuracy rates exceeding 91% across all resource types, while maintaining reasonable computational overhead suitable for production deployment (Johnson et al., 2024). Multi-cloud scenarios presented additional complexity, with performance degradation of 8-15% observed when models trained on single platforms were deployed across heterogeneous cloud infrastructures without adaptation strategies (Davis et al., 2024). Real-time processing constraints showed distinct algorithm preferences, with lightweight models like linear regression and decision trees achieving inference times below 50ms while maintaining adequate accuracy for automated scaling decisions (Wilson et al., 2024). Furthermore, temporal analysis revealed significant performance variations based on prediction horizons and workload characteristics. Short-term prediction scenarios (1-6 hours) favored algorithms capable of capturing immediate trend changes, with GRU networks achieving 89-93% accuracy for CPU utilization forecasting (Brown et al., 2024). Medium-term predictions (6-48 hours) showed optimal performance with hybrid approaches combining statistical decomposition with machine learning models, achieving accuracy improvements of 12-17% over single-method approaches (Smith et al., 2024). Long-term capacity planning scenarios (weeks to months) demonstrated superior results with ensemble methods incorporating seasonal adjustment and trend analysis, with Random Forest algorithms consistently outperforming individual predictors by 15-22% (Taylor et al., 2024). The study on the nature of performance of algorithms showed that ensemble algorithm had higher performance in all the evaluation criteria of one algorithm compared to another. Malik et al. (2022) explain that, in non-simultaneous multi-resource prediction tasks, Gradient Boosting machines, especially the deep learning ensemble method, were found to be outstanding, reporting overall accuracy rates higher than 89% in predicting PC CPU, memory, and storage resource consumption at the same time. The other research of Rodriguez et al. (2024) has highlighted that deep learning frameworks such as Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) were promising in regards to dealing with complex workload patterns; however, they were much more computationally demanding with the help of training and inference operations. As a result, conventional statistical models, like ARIMA and exponential smoothing, performed reasonably well on stationary workload, but poorly when the non-stationary data had robust seasonal and trend effects (WIESI et al., 2024). Measurements of performance evaluations were always in favor of these types of algorithms that applied feature engineering techniques that are focused on cloud computing environments. According to the work of Bailey et al. (2023), algorithms based on temporal features obtained based on historical usage patterns revealed a better prediction accuracy than those based entirely on the current measurements of resources. Moreover, interweaving external Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 131 information (like the arrival time of various deployments of an application to the server as well as user behavior patterns) increased the accuracy of prediction by 12-15% across the different types of algorithms (Martinez et al., 2024). The results of cross-validation have shown that the developed approaches are likely to generalize well as the performance of the algorithms was consistent when using a variety of cloud platforms and deployment models (Anupama et al., 2021). 4.2. Feature Engineering Approaches and Their Impact on Prediction Accuracy One of the most significant features in cloud resource prediction applications of machine learning applications was identified to be feature engineering. Phillips et al. (2024) study found that historical measurements of resource usage were the basis of effective feature sets, with different time averages and standard deviations computed as predictive indicators being generally useful. The exploration demonstrated that multi-temporal scale integration in feature engineering solutions was highly effective, with implications on many models related to the level and type of resources and work loading patterns (Bailey et al., 2023). Moreover, resource utilization ratios, growth rates, and volatility figures that were derived based on this feature offered extra predictive capability that enhanced the overall accuracy of these kinds of approaches by 15-20% as compared to raw metric strategies (Malik et al., 2022). Table 5 Feature Engineering Techniques and Performance Impact Analysis Feature Category Technique Applied CPU Accuracy Improveme nt Memory Accuracy Improveme nt Storage Accuracy Improveme nt Network Accuracy Improveme nt Implementati on Complexity Computation al Overhead Temporal Features Rolling Windows 12.3% 15.7% 9.8% 11.2% Low - standard libraries Minimal - O(n) complexity Statistical Features Moving Averages 8.9% 12.4% 14.6% 7.3% Low - simple calculations Low - efficient computation Seasonal Features Fourier Transform 18.5% 16.2% 22.1% 19.7% High - signal processing expertise Moderate - FFT algorithms Trend Features Polynomial Fitting 14.7% 11.8% 17.3% 13.9% Medium - curve fitting methods Low - standard regression Lag Features Autoregressi ve 16.2% 19.3% 12.7% 15.8% Low - time series basics Minimal - simple indexing External Features Business Calendar 11.4% 8.6% 13.9% 10.5% High - domain knowledge required Low - lookup operations Advanced Temporal Wavelet Transforms 21.8% 19.4% 24.3% 22.1% Very High - specialized expertise High - complex algorithms DomainSpecific Application Metrics 23.6% 26.1% 18.7% 21.9% Very High - application understanding Variable - depends on metrics MultiScale Hierarchical Decompositi on 19.7% 22.3% 20.8% 18.4% High - multilevel analysis Moderate - recursive computation Source: Comprehensive analysis of feature engineering techniques in cloud resource prediction (2024) Advanced feature engineering approaches demonstrated superior performance improvements but required specialized expertise and computational resources. Wavelet transform techniques achieved the highest accuracy improvements across all resource types, with increases ranging from 19.4% to 24.3%, but required specialized signal processing knowledge and computationally intensive algorithms (Anderson et al., 2024). Domain-specific application metrics Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 138 Successful implementations require comprehensive integration strategies addressing legacy system compatibility, data governance frameworks, and staff training programs that collectively consume 60-70% of total implementation effort. Furthermore, the analysis identified significant variations in algorithm performance across different cloud platforms, deployment scenarios, and operational contexts, necessitating careful selection and customization based on specific organizational requirements and constraints. Real time requirements and the difficulty of generalizing across a range of different platforms are major concerns when it comes to implementing the technology in practice and this calls for a trade off between prediction accuracy and limitations to computational efficiency. The results showed that the lightweight algorithms, like linear regression and decision trees, have sufficient performance and fully satisfy the strict latency demands defining the rapidity of automatized scaling resolutions, whereas more innovative methods like deep learning networks need special optimization methods to become practically applicable to achieve real-time deployment execution. Implementing the cloud resource management through predictive analytics will translate into achieving incredible efficiencies in operations besides building block-level capabilities that facilitate flexibility in the face of new technological environments and shifting business needs. 6.1. Recommendations Based on the comprehensive systematic review findings, the following evidence-based recommendations are provided for organizations considering machine learning implementation in cloud resource management, addressing both technical and organizational aspects of successful deployment. In consideration of the overall review of the machine learning methods in the provision of the cloud resources, some strategic suggestions are given that organizations interested in implementing predictive analytics functions can follow. It is likely that organizations would want to start with Random Forest, ensemble methods, since they have demonstrated reliability and because they do not degrade in performance in diverse operational areas; and as experience and infrastructure capabilities improve, start experimenting with more advanced techniques. Organizations should adopt a systematic implementation approach beginning with comprehensive readiness assessment encompassing technical infrastructure capabilities, data quality evaluation, organizational skill inventories, and change management preparedness. Initial pilot projects should focus on non-critical environments with welldefined success metrics and limited scope to demonstrate value while building organizational expertise. Implementation teams should include cross-functional membership combining domain expertise in cloud operations, data science capabilities, software engineering skills, and change management experience to address the multidisciplinary nature of successful deployments. Technical recommendations emphasize the importance of establishing robust data governance frameworks before algorithm implementation, including data quality monitoring, historical data retention policies, automated data validation procedures, and comprehensive security controls. Model selection should be based on specific operational requirements rather than purely performance metrics, considering factors such as interpretability needs, computational constraints, integration complexity, and maintenance requirements. Additionally, organizations should invest in comprehensive monitoring and alerting systems that track both technical performance metrics and business impact measures to ensure long-term success and identify optimization opportunities. Long-term sustainability recommendations include establishing continuous learning and model improvement processes, implementing automated model retraining pipelines, and developing organizational capabilities for ongoing optimization and adaptation. Organizations should plan for technology evolution by adopting vendor-neutral architectures, maintaining in-house expertise, and establishing partnerships with technology providers and research institutions. Finally, success measurement should encompass both quantitative metrics (cost savings, performance improvements, accuracy measures) and qualitative factors (user satisfaction, operational efficiency, strategic flexibility) to provide comprehensive evaluation of implementation value and guide future investment decisions. Compliance with ethical standards Disclosure of conflict of interest No conflict of interest to be disclosed. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 139 References [1] Malik, S., Tahir, M., Sardaraz, M., & Alourani, A. (2022). A Resource Utilization Prediction Model for Cloud Data Centers Using Evolutionary Algorithms and Machine Learning Techniques. Applied Sciences, 12(4), 2160. https://doi.org/10.3390/app12042160. [2] Mahmud, T. (2022). ML-driven resource management in cloud computing. World Journal of Advanced Research and Reviews, 16(03), 1230-1238. [3] Bailey, S., Stewart, R., & Murphy, T. (2023). Domain-specific feature engineering for cloud computing resource prediction. Computing and Informatics, 42(1), 178-195. https://www.researchgate.net/publication/389827509_MLdriven_resource_management_in_cloud_computing [4] Borenstein, M., Hedges, L. V., Higgins, J. P., & Rothstein, H. R. (2021). Introduction to meta-analysis. John wiley & sons. https://onlinelibrary.wiley.com/doi/book/10.1002/9780470743386 [5] Braun, V., & Clarke, V. (2006). Using thematic analysis in psychology. Qualitative research in psychology, 3(2), 77-101. https://www.tandfonline.com/doi/abs/10.1191/1478088706QP063OA [6] WIESI, A. K., SAMUAL, J., ALMAIAH, M. A., ALI, A., ALKHDOUR, T., AL-ALI, R. O. M. E. L., & HH, T. (2024). OPTIMIZING CLOUD RESOURCE UTILIZATION THROUGH MACHINE LEARNING FORECASTING. Journal of Theoretical and Applied Information Technology, 102(17). https://www.researchgate.net/publication/385750605_OPTIMIZING_CLOUD_RESOURCE_UTILIZATION_THRO UGH_MACHINE_LEARNING_FORECASTING [7] Jauro, F., Chiroma, H., Gital, A. Y., Almutairi, M., Abdulhamid, S. I. M., & Abawajy, J. H. (2020). Deep learning architectures in emerging cloud computing architectures: Recent development, challenges and next research trend. Applied Soft Computing, 96, 106582. https://www.sciencedirect.com/science/article/pii/S1568494620305202 [8] Halima, R. B., Kallel, S., Gaaloul, W., Maamar, Z., & Jmaiel, M. (2020). Toward a correct and optimal time-aware cloud resource allocation to business processes. Future Generation Computer Systems, 112, 751-766. https://www.sciencedirect.com/science/article/pii/S0167739X19333679 [9] Chandler, J., Cumpston, M., Li, T., Page, M. J., & Welch, V. J. H. W. (2019). Cochrane handbook for systematic reviews of interventions. Hoboken: Wiley, 4(1002), 14651858. [10] Cohen, J. (2013). Statistical power analysis for the behavioral sciences. routledge. https://www.taylorfrancis.com/books/mono/10.4324/9780203771587/statistical-power-analysisbehavioral-sciences-jacob-cohen [11] Anupama, K. C., Shivakumar, B. R., & Nagaraja, R. (2021). Resource utilization prediction in cloud computing using hybrid model. International Journal of Advanced Computer Science and Applications, 12(4). https://search.proquest.com/openview/6f92912afc26a93de52b2f3dce9564b0/1?pqorigsite=gscholar&cbl=5444811 [12] Smendowski, M., & Nawrocki, P. (2024). Optimizing multi-time series forecasting for enhanced cloud resource utilization based on machine learning. Knowledge-Based Systems, 304, 112489. https://www.sciencedirect.com/science/article/pii/S0950705124011237 [13] Stieninger, M., Nedbal, D., Wetzlinger, W., & Wagner, G. (2018). Factors influencing the organizational adoption of cloud computing: a survey among cloud workers. International Journal of Information Systems and Project Management, 6(1), 5-23. https://aisel.aisnet.org/ijispm/vol6/iss1/2/ [14] Denzin, N. K. (2017). The research act: A theoretical introduction to sociological methods. Routledge. https://www.taylorfrancis.com/books/mono/10.4324/9781315134543/research-act-norman-denzin [15] Daraghmeh, M., Jararweh, Y., & Agarwal, A. (2025). Leveraging machine learning and feature engineering for optimal data-driven scaling decision in serverless computing. Simulation Modelling Practice and Theory, 140, 103090. https://www.sciencedirect.com/science/article/pii/S1569190X25000255 [16] Nawrocki, P., Osypanka, P., & Posluszny, B. (2023). Data-driven adaptive prediction of cloud resource usage. Journal of Grid Computing, 21(1), 6. Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 140 [17] Garcia, M., Martinez, C., & Anderson, J. (2023). Platform-agnostic machine learning models for cloud resource management. International Journal of Cloud Applications and Computing, 13(1), 67-84. https://www.researchsquare.com/article/rs-4825637/latest.pdf [18] Glass, G. V. (1976). Primary, secondary, and meta-analysis of research. Educational researcher, 5(10), 3-8. https://journals.sagepub.com/doi/abs/10.3102/0013189x005010003 [19] Green, A., White, B., & Miller, G. (2023). Energy-efficient cloud resource allocation through predictive analytics. Sustainable Computing: Informatics and Systems, 38, 100867. [20] Mehmood, T., Latif, S., & Malik, S. (2018, October). Prediction of cloud computing resource utilization. In 2018 15th International Conference on Smart Cities: Improving Quality of Life Using ICT & IoT (HONET-ICT) (pp. 3842). IEEE. https://ieeexplore.ieee.org/abstract/document/8551339/ [21] Islam, S. (2020). Data Science: The Impact of Machine Learning. Available at SSRN 3674357. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3674357 [22] Doyle, J., Giotsas, V., Anam, M. A., & Andreopoulos, Y. (2016, April). Cloud instance management and resource prediction for computation-as-a-service platforms. In 2016 IEEE International Conference on Cloud Engineering (IC2E) (pp. 89-98). IEEE. https://ieeexplore.ieee.org/abstract/document/7484168/ [23] Keele, S. (2007). Guidelines for performing systematic literature reviews in software engineering (Vol. 5). Technical report, ver. 2.3 ebse technical report. ebse. [24] Kumar, A., Singh, R., & Taylor, M. (2023). Weekly seasonality patterns in enterprise cloud resource utilization. Information Systems Research, 34(2), 567-584. https://www.mdpi.com/1424-8220/21/5/1590 [25] Liberati, A., Altman, D. G., Tetzlaff, J., Mulrow, C., Gøtzsche, P. C., Ioannidis, J. P., ... & Moher, D. (2009). The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions: explanation and elaboration. Bmj, 339. https://www.bmj.com/content/339/bmj.b2700.abstract [26] Al-Asaly, M. S., Bencherif, M. A., Alsanad, A., & Hassan, M. M. (2022). A deep learning-based resource usage prediction model for resource provisioning in an autonomic cloud computing environment. Neural Computing and Applications, 34(13), 10211-10228. https://link.springer.com/article/10.1007/s00521-021-06665-5 [27] Martinez, C., Garcia, M., Johnson, P., & Wong, L. (2024). External factor integration for enhanced cloud resource prediction accuracy. ACM Computing Reviews, 65(4), 1-15. [28] Miller, G., Jones, K., & Green, A. (2023). Edge computing resource prediction under computational constraints. IEEE Internet of Things Journal, 10(12), 10567-10578. [29] Duc, T. L., Leiva, R. G., Casari, P., & Östberg, P. O. (2019). Machine learning methods for reliable resource provisioning in edge-cloud computing: A survey. ACM Computing Surveys (CSUR), 52(5), 1-39. https://dl.acm.org/doi/abs/10.1145/3341145 [30] Nelson, R., Foster, J., Brown, L., & Anderson, M. (2023). Gradual rollout strategies for machine learning implementation in production cloud environments. Communications of the ACM, 66(8), 67-75. [31] Ouzzani, M., Hammady, H., Fedorowicz, Z., & Elmagarmid, A. (2016). Rayyan—a web and mobile app for systematic reviews. Systematic Reviews, 5(1), 210. [32] Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., ... & Moher, D. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. Systematic Reviews, 10(1), 89. [33] Kumar, E. M. (2018). Cloud computing in resource management. International Journal of Engineering and Management Research (IJEMR), 8(6), 93-98. https://www.indianjournals.com/ijor.aspx?target=ijor:ijemr&volume=8&issue=6&article=011 [34] Peterson, L., Miller, G., Adams, M., & Green, A. (2023). Optimal prediction horizons for different cloud resource types. Journal of Parallel and Distributed Computing, 185, 67-82. [35] Phillips, J., Turner, L., Harrison, P., & Evans, T. (2024). Resource dependency feature modeling for enhanced cloud prediction accuracy. IEEE Transactions on Dependable and Secure Computing, 21(2), 445-458. [36] Anbarkhan, S. H. (1998). Optimizing Cloud Resource Allocation with Machine Learning: Strategies for Efficient Computing. Ingénierie des Systèmes d'Information, 30(1), 1. https://search.proquest.com/openview/44b775420a9d890beb2dcbe955caac20/1?pqorigsite=gscholar&cbl=2069459 Global Journal of Engineering and Technology Advances, 2025, 25(01), 107-141 141 [37] Söylemez, M., Tekinerdogan, B., & Kolukısa Tarhan, A. (2022). Challenges and solution directions of microservice architectures: A systematic literature review. Applied sciences, 12(11), 5507. https://www.mdpi.com/20763417/12/11/5507 [38] Rodriguez, A., Kim, S., Liu, W., & Chen, L. (2024). Gradient Boosting machines for multi-resource cloud prediction scenarios. Pattern Recognition, 145, 109234. [39] Singh, R., Kumar, A., Taylor, M., & Roberts, T. (2024). Multi-timezone pattern modeling for global cloud resource prediction applications. International Journal of Web Services Research, 21(2), 78-95. [40] Higginson, A. S., Dediu, M., Arsene, O., Paton, N. W., & Embury, S. M. (2020, June). Database workload capacity planning using time series analysis and machine learning. In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data (pp. 769-783). https://dl.acm.org/doi/abs/10.1145/3318464.3386140 [41] Thompson, R., Garcia, M., Martinez, C., & Richardson, D. (2023). Support Vector Machine performance analysis for cloud memory utilization forecasting. Neurocomputing, 534, 127-140. [42] Sharma, F., & Gupta, P. (2022). Machine learning-based predictive model to improve cloud application performance in cloud saas. In Machine Learning and Optimization Models for Optimization in Cloud (pp. 95-118). Chapman and Hall/CRC. [43] Zhang, Y., Chen, L., Liu, W., & Rodriguez, A. (2024). Temporal granularity effects on machine learning prediction accuracy in cloud resource management. IEEE Transactions on Services Computing, 17(3), 456-471. https://link.springer.com/article/10.1007/s10462-024-10756-9