Call Availability Prediction in a Telecommunication System: A Data Driven Empirical Approach

Abstract

Availability prediction in a telecommunication system plays a crucial role in its management, either by alerting the operator to potential failures or by proactively initiating preventive measures. In this paper, we apply linear (ARMA, multivariate, random walk) and nonlinear (Radial and Universal Basis Functions) regression techniques to recognize system failures and to predict the system's call availability up to 15 minutes in advance. Secondly we introduce a novel nonlinear modeling technique for call availability prediction. We benchmark all five techniques against each other. The applied modeling methods are data driven rather than analytical and can handle large amounts of data. We apply the modeling techniques to real data of a commercial telecommunication platform. The data used for modeling includes: a) time stamped event-based log files; and b) continuously measured system states. Results are given in terms of a) receiver operator characteristics (AUC) for classification into classes of failure and non-failure states and b) as a cost-benefit analysis. Our findings suggest: a) high degree of nonlinearity in the data; b) statistically significant improved forecasting performance and cost-benefit ratio of nonlinear modeling techniques; and finally finding that c) log file data does not contribute to improve model performance with any modeling technique