Pattern analysis using event-covering
David Chiu · 1986
Computer-based pattern analysis on a set of observations can acquire useful knowledge if the methodology is domain independent and applicable to complex tasks. When there exists a certain degree of uncertainty because of probabilistic variations and noise perturbation, the probabilistic approach is then very useful in the formulation of a knowledge acquisition and reasoning scheme. This dissertation proposes a probabilistic (information-theoretic) event-covering approach for pattern analysis tasks which are applicable to data of the general type--including: (1) the incomplete multivariate discrete-valued (symbolic) data, (2) the mixed discrete and continuous valued data and (3) time-dependent discrete-valued data. The event-covering approach detects statistically significant event associations and can deduce a certain structure of inherent data relationships. Event-covering refers to the process of covering or selecting statistically significant outcomes in the outcome space of variable-pairs for a pattern analysis task, disregarding whether the variables considering the complete outcome space as a whole is statistically significant or not. When the statistically significant information obtained from the event-covering process is synthesized, an effective reasoning method can be formulated. The pattern analysis tasks included in this dissertation are: probabilisitic inference, supervised classification, cluster analysis and forecasting. The developed methods have been extensively evaluated in experiments using simulated data with different dependency structure and real life data for empirical knowledge extraction.