Improving Database Management System Efficiency Through Data Mining Approaches

Cloud computing has become increasingly popular owing to its economic advantages, elastic scalability, and flexible pay-per-use pricing model, a trend further driven by the rapid expansion of Internet of Things(IoT) applications. This paper focuses on improving the performance of shared cloud-based databases operating under combined transactional and analytical workloads by introducing two complementary techniques. First, business intelligence and decision-support queries frequently issue complex aggregate sub-expressions. A scalable tree-mining approach is used to discover commonly recurring sub-expressions across historical query execution plans. These sub-expressions are then materialised as views, reducing redundant computation. Second, web-based applications such as e-commerce platforms and online banking systems exhibit time-of-day and location-specific access patterns. A predictive framework based on Parzen window estimation identifies evolving working sets. Building on this, two cache replacement policies are proposed: Predictive LRLFU (Pr.LRLFU) with fixed beta, and Adaptive LRLFU (Ad.LRLFU) which tunes beta online through a ghost-list feedback loop. Experimental results show that the proposed techniques achieve notable improvements in cache hit ratios and overall system performance compared with LRU, LFU, LIRS, ARC, and CAR. The accompanying reference implementation is available at https://github.com/DagurKapil/thesis_src. Status: Preprint. Submitted for peer review.