Adaptive Machine Learning-Driven Enhancements for TREEPROMPT and TREEHP2PL in Distributed Real-Time Database Systems

Abstract

Managing nested transactions in distributed real-time database systems (DRTDBS) is essential for ensuring consistency, scalability, and efficiency in critical domains such as financial systems and industrial automation. While traditional protocols like TREEPROMPT resolves inter-transaction deadlocks with speculative execution and priority inheritance and TREEHP2PL detects intra-transaction deadlocks using Wait-For Graphs and resolves them by aborting low-priority or short-execution subtransactions, their static configurations limit adaptability to dynamic workloads. This study enhances these protocols by integrating machine learning (ML) classification models to improve performance through predictive success analysis. Four ML models—Naive Bayes, Decision Tree, K-Nearest Neighbours (KNN), and Random Forest—are evaluated using a dataset of 500 transactions per simulation run, with ten independent executions to ensure statistical reliability. The experimental setup evaluates classifier performance using accuracy, precision, recall, F1-score, and computational efficiency. The results show that the Naive Bayes model achieved an accuracy of 98.5%, a precision of 98.2%, a recall of 98.7%, and an F1-score of 98.4%. The Decision Tree model performed similarly, with an accuracy of 97.8%, a precision of 97.5%, a recall of 97.9%, and an F1-score of 97.6%. In contrast, the K-Nearest Neighbours (KNN) model exhibited lower performance, with an accuracy of 44.2%, a precision of 43.8%, a recall of 44.5%, and an F1-score of 44.1%. Similarly, the Random Forest model achieved an accuracy of 45.6%, a precision of 45.3%, a recall of 45.9%, and an F1-score of 45.5%. Compared to traditional heuristic-based approaches, ML-enhanced protocols significantly improve transaction success rates by minimizing deadlock occurrences and optimizing resource utilization. Moreover, ML integration enhances system throughput and reduces transaction latency, demonstrating notable computational efficiency gains. These findings validate the effectiveness of ML-driven optimizations in enhancing protocol scalability and adaptability. Future research will focus on refining underperforming models, incorporating reinforcement learning techniques, and testing on larger datasets to further optimize real-time transaction management in DRTDBS environments.