Pattern Recognition for Hidden Markov Processes: Locality and Accuracy

Hidden Markov models (HMMs) have been used as probabilistic models for important tasks such as speech recognition, natural language processing, and process mining. A hidden Markov process is a workflow viewed as a HMM with transition arcs connecting the operational stages of the workflow as a directed graph. In a complex environment, multiple processes may run concurrently by various agents. By extracting event sequences associated with the corresponding processes from an event log, we can identify the operational traces of the underlying processes. An important pattern recognition task for process management is to analyze these event sequences and match them with the most likely workflow models. This enables the system to recognize the operational intentions of the agents for resource management and anomaly detection. In this paper, we investigate locality structures embedded in the transition probabilities and the observation probabilities of hidden Markov processes. We parameterize the locality structures and conduct an empirical study under various parameter settings in the context of text keyboarding by people with special needs using a virtual keyboard. Our empirical results reveal how the locality structures directly affect the predictive accuracy attained for the process recognition task. The findings provide insight into the feasible ranges in the locality parameter space for effective process recognition.