基于自编码网络的移动轨迹异常检测

Anomaly Detection in Mobile Trajectory Using Auto-encoder Network

移动轨迹异常检测是指从一群轨迹中寻找偏离一般模式的轨迹。基于聚类的异常检测依赖成对轨迹的距离计算,其计算实时性差且异常检测准确率低。提出面向移动轨迹异常检测的自编码网络,该模型对一般模式的轨迹有鲁棒的向量化表达能力,能够重构出与原始轨迹相近的输出;而对于偏离一般模式的轨迹敏感,重构后的输出与原始轨迹的差异大。根据该差异可直接检测异常,无需计算轨迹间的距离。以出租车轨迹为研究对象,试验结果表明该模型有更高的F Score,并且在数据量较大时检测时间低于参照方法,因此在高动态、大数据量的场景具有更好的适用性。

Mobile trajectory anomaly detection refers to finding the trajectories deviating from the general pattern. Cluster-based anomaly detection methods rely on the calculation of paired distances, which has high computational complexity. An auto-encoder network for moving trajectory anomaly detection is proposed. The model has a robust vectorized representation of the normal data, and is able to reconstruct an output close to original trajectory. However, the model is sensitive to the data deviating from general patterns, which means that the deviation between the reconstructed output and original trajectory is large. We use the deviation to discriminate whether the trajectory is anomaly, without calculating the paired distances. We conducted experiments using taxi trajectories. The result shows that our method has higher F Score, and the detection time is lower than the reference methods when the amount of data is large. Therefore, our method has better applicability in real-time and large-data-volume scenarios.