Content area
Smartphone applications and IoT devices operate in dynamic environments where context events—such as battery status, screen orientation changes, and network connectivity—directly influence application behavior, reliability, and user experience. The volatility and complexity of these events pose challenges for application development, testing, and performance optimization. This dissertation addresses these challenges by introducing novel machine learning and data mining frameworks to explore, predict, and leverage context event patterns. We employ sequence rule mining (POERMH, TKE-Rules) and predictive models to identify frequent context event sequences, achieving strong Recall, Precision, and F-1 scores—enhancing testing efficiency for common user scenarios. A modified Compact Prediction Tree further improves context data prediction, where All-k Order Markov and Transition Directed Acyclic Graph models outperform baselines in high-frequency event forecasting. High-utility itemset mining (TKQ, FHUQI-Miner, FCHM) is applied to prioritize high-impact testing scenarios, integrating decision tree regression and bagging for enhanced predictive accuracy. Additionally, using CM-SPAM and machine learning models (Random Forest, Support Vector Machine, and Long Short-Term Memory), we achieve high-accuracy application behavior prediction, with Random Forest performing best at one-minute intervals.
By integrating sequence mining, utility-driven pattern discovery, and machine learning, this research significantly improves application testing efficiency and user experience. These approaches address sparsity, volatility, and complexity in context event data, enabling developers to anticipate user behavior, optimize resource usage, and design robust, context-aware applications.