|
Abstract:
|
Large-scale service systems generate a stream of tickets describing customer-affecting events and their remediation. For analytic purposes, we would like to label the tickets as belonging to various classes of interest. For example, in systems handling weather-related service tickets, it is important to isolate the class of storm-related tickets. Automation of the labeling process can greatly enhance the efficiency of decision-making processes in disturbance management, including prediction, monitoring, planning and resource allocation. The resulting labels are typically probabilistic, due to issue complexity, uncertainty, missing information, costs and other factors. In this paper, we describe an approach to probabilistic labeling of service tickets that utilizes the assumed stochastic properties of the underlying disturbance origination and development. The focus of the paper is on use of robust estimation techniques for the baseline ticket rates, model validation and inference related to the established probabilistic labels.
|
