Identifying Mobility of Drug Addicts with Multilevel Spatial-Temporal Convolutional Neural Network

Abstract

Human identification according to their mobility patterns is of great importance for a wide spectrum of spatial-temporal based applications. For example, detecting drug addicts from normal residents in public security area. However, extracting and classifying user behaviors in massive amount of moving records is not trivial because of three challenges: (1) the complex transition records with noisy data; (2) the heterogeneity and sparsity of spatiotemporal trajectory features; and (3) extremely imbalanced data distribution of real world data. In this paper, we propose MST-CNN, a multi-level convolutional neural network with spatial and temporal features. We first embed the multiple factors on single trajectory level and then generate a behavior matrix to capture the user’s mobility patterns. Finally, a CNN module is used to extract various features with different filters and classify user type. We perform experiments on real-life mobility datasets provided by public security office, and extensive evaluation results demonstrate that our method obtains significant improvement performance in identification accuracy and outperform all baseline methods.