Prediction of Workloads in Incident Management Based on Incident Ticket Updating History

摘要

Incident management is one of the most important and burdensome tasks in system management. In order to achieve effective incident management, prediction of the workload needed to solve incidents is quite useful. Using this prediction, we can provide a fair distribution of incident tickets to administrators. In order to predict the workload needed to handle an incident ticket when it arrives, we propose an incident ticket classification method based on text mining (TF-IDF and Naive Bayes). In this approach, we first collect incident tickets with their number of updates as workload indicators. Next, we construct a model representing the relation between the words in incident texts and the incident workload category (easy or difficult) based on Naive Bayes. We then predict a category into which each new incident ticket should be classified using the model. We implemented our method using Hadoop and Mahout library. By conducting the evaluation with incident tickets recorded in an cloud infrastructure for research, we confirmed that our approach can predict the workload of incident tickets with an F-measure of 0.81 in its best case.