PCA-Aware Anomaly Correction for Traffic Matrix in an IP Backbone Network

摘要

Traffic matrices (TM) describe the traffic flows in IP networks, and TM prediction is a critical application in network planning. Evidenced by the TM measurements in a IP backbone network, there exist various kinds of traffic anomalies, which are not caused by malicious attacks but have great negative effects on TM prediction. It is a challenge problem to correct these anomalies without introducing additional prediction errors. To improve the prediction performance against these anomalies, we propose to utilize principle component analysis (PCA) in the data correction phase of TM prediction. A threshold is introduced to indicate the change level of the trend components during TM correction. Once the change is less than this threshold, the variance of nodal traffic time serials is corrected by linear interpolation method in a manner of sliding window. Experiment results show that, this PCA-aware TM correction method can effectively improve the TM prediction performances, either in terms of total traffic or individual flows. When the change of principal component is controlled less than a given threshold (around 3.5% in our case), the effects of traffic anomalies can be eliminated by our method.