Network management systems play an important role to deal with the large size and complexity of current cellular networks. Thus, operators and vendors focus much of their efforts on developing new techniques and tools for network management. One of the most critical processes in network management is fault management, since a failure in a network element might have a strong impact on user satisfaction due to service degradation. Unfortunately, cellular networks generate thousands of alarms daily, which
have to be checked manually by operator personnel. With the latest advances in big data analytics, different methods for reducing the number of alarms to be monitored have been proposed in the literature. In this work, an automatic method for prioritizing alarms based on the need for specialized personnel is presented. The core of the method is an ensemble model built with supervised learning that estimates the probability that an alarm generates a trouble ticket. The model is trained with trouble ticket data from the network operation center. A performance comparison of four classical base classifiers (naïve Bayes, random forest, artificial neural network and support vector machine) for the ensemble is presented. The model is implemented in IBM SPSS Modeler and tested with a real alarm and trouble ticket dataset taken from a live cellular network. Results show that the proposed model correctly flags those alarms that need further analysis by the operator and filter out those alarms that do not have impact on network performance. The main contribution of this work is unveiling a new application (the automatic prioritization of alarms in a cellular network based on the need for specialized personnel) and presenting for the first time a performance comparison of base classifiers used for this purpose (since the required dataset is extremely difficult to find for privacy reasons).
