Modern SLA management includes SLA prediction based on data collected during service operations. Besides overall accuracy of a prediction model, decision makers should be able to measure the reliability of individual predictions before taking important decisions, such as whether to renegotiate an SLA. Measures of reliability of individual predictions provided by machine learning techniques tend to depend strictly on the technique chosen and to neglect the features of the system generating the data used to learn a model, i.e., the service provisioning landscape in this case. In this paper, we define a hybrid measure of reliability of an individual SLA prediction for classification models, which accounts for both the reliability of the chosen prediction technique, if available, and features capturing the variability of the service provisioning scenario. The metric is evaluated empirically using SLAs and event logs of a real world case.
This paper was presented in ACM Symposium on Applied Computing (SAC) in April 2019 (GGS Class 2).