A streaming approach to reveal crowded events from cellular data

Anomaly detection has been a very popular research topics over the last few years and applies to many scenarios from different disciplines. This research focuses on crowded phenomena, and addresses the detection of popular events by looking for “anomalous” patterns in cellular traffic data. In particular, the paper elaborates upon previous proposals and presents two streaming algorithms based on the wavelet decomposition of traffic data. The new algorithms consume traffic samples as soon as they are available, elaborate the data in real time and possibly raise alarms upon threshold crossing. The effectiveness of the approach is assessed by using the public dataset containing the real cellular data acquired over the network of the most popular Italian traffic operator. The experiments prove that the streaming algorithms generally achieve performance comparable to that of their offline counterparts, and that the small degradation that may occasionally be observed is however well counterbalanced by the obvious advantage of detecting anomalies in real-time with no need to wait for the elaboration of overly long traffic timeseries.