Mining Predictive Patterns in Sequences of Events

摘要: Learning to predict rare events from sequences of with categorical features is an important, real-world, problem. Unfortunately, most machine learning methods that learn classification “rules” are not suited solving this type problem because they assume unordered set examples and cannot identify patterns between “examples” (i.e., events). Statistical time-series prediction also suitable, since numerical features. Genetic algorithms, however, which have often been used find in data, well finding predictive temporal sequential the event sequence data. In order solve problem, we developed Timeweaver, a genetic-based system that, given pre-specified “target” event, learns data successfully future occurrence event. Timeweaver has applied task predicting telecommunication failures time-stamped alarm messages outperformed several simple methods. The described KDD paper (Weiss & Hirsh 1998), as GECCO provides more detailed description GA 1999). Due availability these papers, only short provided workshop notes. We employ Michigan-style evolve rules. main issue faced when applying evolutionary algorithm was balance exploration search space efficiency search. particular, needed appropriate fitness function diversity maintenance strategy. Our factored both recall each rule (the percentage target predicted) precision correct predictions). way two measures were combined found dramatic impact on strategy adopted involves varying importance measures, so population evolved contains some highly precise rules less cover events. A niching called sharing ensure diverse developed, collectively majority