Complex information systems generate large amount of event logs that represent the states of system dynamics. By monitoring these logs, we can learn the process models that describe the underlying business procedures, predict the future development of the systems, and check whether the process models match the expected ones. Most of the existing process monitoring techniques are derived from the workflow management systems used to cope with the logs generated by systems with deterministic outcomes. In this dissertation, however, I consider novel techniques that handle event log data, monitor system deviations, and infer the development of systems based on probabilistic process models. In particular, I present a novel process monitoring approach based on maximizing the information divergences of the system state dynamics and demonstrate its efficiency in detecting abrupt changes, as well as long-term system deviation. In addition, a new process modeling technique, Classification Tree hidden (semi-) Markov Model (CTHMM), is proposed. I show that CTHMM derived from Classification and Regression Tree and hidden semi-Markov model (HSMM) with hidden system states identified by Classification Tree can help discover and predict relevant system state sequences in temporal-probabilistic manners. The main contributions of this dissertation can be summarized as follows: 1) a new approach used in process monitoring that helps detect anomalies of dynamic systems from the point of views of both system change-point and long-term system deviation; 2) a unique HMM/HSMM learning technique that solves the problem of hidden state splitting and estimates HMM/HSMM parameters simultaneously; 3) a novel temporal-probabilistic process model that generates human-comprehensible IF-THEN system state definitions used to help infer evolutions of discrete dynamic systems.
展开▼
