S2N2: An interpretive semantic structure attention neural network for trajectory classification

• Main Authors: JIN, Canghong, TAO, Ting, LUO, Xianzhe, LIU, Zemin, WU, Minghui
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2020
• Subjects: human behavior understanding; interpretive trajectory structure; User classification; Software Engineering
• Online Access: https://ink.library.smu.edu.sg/sis_research/5143; https://ink.library.smu.edu.sg/context/sis_research/article/6146/viewcontent/09044862_pvoa.pdf
• Institution: Singapore Management University

We have witnessed a rapid growth over past decades in sensor data mining (SDM), which aims at extracting valuable information automatically from large repositories of moving activity data. One of the significant SDM tasks is identifying humans through their transit modes using a variety of user-tracking systems. However, to the best of our knowledge, distinguishing traces of users and understanding their behaviors are difficult tasks in most real-life cases for the following reasons: 1) activity data containing both temporal and spatial contexts are of high order and sparse; 2) living patterns are not as regular as expected, and the route choice uncertainties due to their vagueness and randomness; 3) owing to the complexity and sparseness of urban travel methods, although some deep learning-based models can produce relatively good classification results, they can still be improved by combining external information. To address these challenges, we propose a novel scenario-based deep learning method which is based on the assumption that people visit places with explicit purposes (e.g., to go to work or visit a park). We first represent semantic patterns from daily life and create various scenarios and utilize an attention neural network to embed points of trajectories by considering both semantic and geographical information. Then, we construct a Semantic Structure Neural Network (S2N2) framework to perform the end-end classification. Our S2N2 model is applied to an interesting yet challenging topic: distinguishing suspect transit behavior on a real-life data set collected by mobile devices. Although the problem is not entirely solved, the extensive evaluation presented here demonstrates that our model outperforms conventional classification methods, anomaly detection methods, and state-of-the-art sequential deep learning models, especially when trajectory semantic vectors are incorporated. We also provide statistical analysis and intuitive explanations to help interpret the characteristics of user mobility.