Mobility-Based Anomaly Detection in CDR Based Trajectories from the Mobile Cellular Network

In the last decade, it has been proven that using mobile data or Call Detail Records can be a good data source for understanding human mobility, daily commuting, and population displacement. However, the quality of trajectories extracted from these data can sometimes be inaccurate or erroneous due to dif-ferent anomalies inflected by human error or network hardware. Hence, this paper proposes a model to identify the anomalies in CDR-based trajectories. The created model is designed by relying on a mixed-feature sequence approach that has a preprocessing step for creating a cell area coverage-based profile that is used by an autoencoder for reducing dimensionality, therefore reducing computation needs in the last step, which is a clustering layer that detects the anomalies. The evaluation was conducted in a lab setup using real dataset and the performance was around 70% of accurate detection including the affected cell areas or spotted, which is very encouraging with respect to complexity and origin of the problem that is due to human error and mobile cellular network integrity.