Temporal Pattern Mining in Dynamic Environments

In this work an approach is presented which applies unsupervised symbolic learning to a qualitative abstraction of dynamic scenes in order to create frequent temporal patterns and prediction rules. Having in mind rather complex situations with different objects of various types and relations and temporal interrelations of actions and events, the approach provides means to mine complex temporal patterns taking into account these aspects. It is an extension of the association rule mining algorithm Apriori and combines ideas from relational as well as sequential association rule mining approaches. Temporal interrelations between predicates of patterns are represented qualitatively by interval relations as, e.g., introduced by Allen and Freksa. Additionally, variable unification allows to connect variables of (different) predicates in a complex pattern in order to deal with relational data. As a third aspect, concept restrictions are learned for variables of a pattern.