On-line Network Resource Consumption Prediction with Confidence

Traffic prediction is critically important for network resource management and performance evaluation. Accurate and fast prediction requires algorithmic capability, in particular, machine learning algorithms. Various learning and prediction methods have been developed and applied to provide such capability. However, these methods can only provide bare predictions, i.e. algorithms predicting values for new examples without saying how reliable these predictions are. In this paper, an on-line learning algorithm based on ridge regression is described. The on-line algorithm can give reasonably tight tolerance intervals for regression estimates. The predicted results of the algorithm on two real network traffic datasets show good performance.