Comparative Analysis of Machine Learning Algorithms for Real-Time IoT Data Processing

Thesis Abstract

The rapid proliferation of Internet of Things (IoT) devices has resulted in an exponential increase in real-time data generation, necessitating efficient and accurate data processing mechanisms to support timely decision-making across diverse applications such as healthcare, smart cities, and industrial automation. Despite the growing adoption of machine learning (ML) algorithms for analyzing IoT data streams, there remains a lack of comprehensive comparative assessments to identify the most suitable algorithms in real-time processing contexts, considering factors such as computational efficiency, accuracy, scalability, and adaptability. This study aims to systematically evaluate and compare prominent machine learning algorithms—namely Decision Trees, Random Forests, Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Neural Networks—specifically focusing on their performance in real-time IoT data environments. The primary objectives are to (1) measure the processing speed, classification accuracy, and resource consumption of each algorithm when applied to real-time IoT datasets; (2) analyze their scalability with increasing data volumes; (3) assess their robustness against noisy and incomplete data; and (4) develop a framework to guide optimal algorithm selection for specific IoT applications. The research employs a quantitative approach within a comparative research design, utilizing experimental simulation of IoT data streams collected from a network of 500 sensor nodes deployed in a smart city scenario. Data collection instruments include a custom-developed IoT data emulator capable of generating real-time, multi-modal data streams, supplemented by real-world data obtained from a publicly available urban sensor dataset comprising approximately 2 million data points. Data analysis involves applying descriptive statistics to summarize performance metrics such as latency, accuracy, memory utilization, and energy consumption. Inferential statistical techniques, specifically analysis of variance (ANOVA), are employed to determine statistically significant differences in the performance of different algorithms across various data conditions. Furthermore, machine learning model evaluation incorporates cross-validation methods, and hyperparameter tuning is performed via grid search to optimize each algorithm’s performance. Hierarchical regression analysis is utilized to examine the influence of data noise and size on algorithm robustness, while a multi-criteria decision analysis (MCDA) framework consolidates the multiple performance indicators into an overall ranking of algorithm suitability. Expected findings are anticipated to reveal a trade-off between accuracy and computational efficiency, with Random Forests and Neural Networks demonstrating superior predictive performance but at higher resource costs, whereas simpler models like Decision Trees and k-NN exhibit faster processing times but lower accuracy under complex data conditions. The analysis aims to establish clear performance benchmarks for each algorithm and recommend context-dependent choices aligned with application-specific constraints such as latency requirements and device capabilities. The study also contains a proposed conceptual framework, integrating these performance insights, to assist IoT system architects in selecting optimal machine learning algorithms tailored to real-time data applications. This research substantially contributes to the knowledge base by providing a systematic comparison of ML algorithms in the specific context of real-time IoT data processing, filling existing gaps in literature related to operational efficiency and robustness. The findings can inform practitioners and researchers on deploying appropriate algorithms in resource-constrained environments, thereby enhancing IoT system performance and reliability. The study concludes by emphasizing the importance of context-aware algorithm selection and advocates for further research into hybrid models and adaptive learning systems that dynamically optimize performance based on evolving data patterns. Recommendations include developing standardized benchmarking protocols for real-time IoT data analytics and extending comparative analyses to emerging deep learning models, such as convolutional and recurrent neural networks, within the IoT domain.