Anomalous Trajectory Detection in Vehicular Social Networks With Unsupervised Dynamic Network Representation Learning

As data volumes surge, the automatic detection of anomalous trajectories has become increasingly difficult in real- world scenarios. However, most existing works fail to account for the spatial-temporal properties of trajectory simultaneously and rely on biased real-world or synthetic datasets. To tackle these challenges, we propose a novel unsupervised method, Dynamic Network Representation learning (DNR), that leverages the advances in the network field and adapts them to transportation. DNR constructs a dynamic network from the perspective of vehicle social networks using the idea of the k-nearest neighbors. A dynamic walk algorithm is proposed to sample nodes and context over time, focusing on topology changes. DNR embeds the network using Skip-Gram network representation and incremental learning to obtain two trajectory embeddings for two anomaly types. Our experiments on four real-world datasets and eight generative datasets demonstrate the effectiveness of DNR over other methods and reveal its rationality to detect deviations from typical trajectories.