Learning Time-based Rules for Prediction of Alarms from Telecom Alarm Data Using Ant Colony Optimization

This paper proposes a new method to learn time based rules from telecom system alarm data for prediction of the classes of alarms. A time based rule associates an alarm class with the StartTime attribute and other attributes of alarms. The rules are evaluated with the coverage of the rules in the training data set. Given a new alarm generated at a particular time, its alarm class can be predicted with a set of time based rules. We present a new algorithm that extracts time based rules from alarm data through an ant colony optimization (ACO) process. Given an alarm training data, a search space is formulated as a square matrix indexed by distinctive attribute values. The pheromone at the search space is computed from the training data and a time based rule is discovered from the pheromone distribution. The pheromone distribution is updated after a time based rule is extracted and the search for a new rule starts. A rule pruning process is used to remove redundant rules and increase the prediction accuracy of the final rule set. We experimented the new method on Nokia Simmons (NSN) and Ericsson data sets and compared the results of the new method and the TimeSeluth system. The comparison demonstrated that the new method outperformed TimeSeluth in prediction accuracy.