Deep learning approach for trajectory user-linking in multidimensional and imbalanced datasets

• Ano de defesa: 2024
• Autor(a) principal: Freitas, Nicksson Ckayo Arrais de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Palavras-chave em Português: Vinculação de usuários a trajetórias; Classificação de trajetórias; Aprendizado profundo; Aprendizado de máquina; Trajetória esparsa; Conjunto de dados desbalanceado; Dados geoespeciais; Trajectory user-linking; Trajectory classification; Deep learning; Machine learning; Sparse trajectory; Imbalanced dataset; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.ufc.br/handle/riufc/79074

In this work, we investigate the trajectory classification problem and focus on the Trajectory User-Linking (TUL) challenge within Location-Based Social Networking (LBSN) to associate anonymous subtrajectories with specific users on platforms such as Foursquare. This association is crucial for enhancing service personalization and targeting, optimizing urban and business planning, and facilitating effective public health and safety strategies. The TUL problem presents multiple challenges: LBSN databases often contain large volumes of data; there is complexity in representing spatial and temporal dimensions in machine learning models; spatiotemporal trajectory points are sparse; LBSN trajectories are multidimensional, that is, there are other features available that are associated with trajectory points; the number of classes often exceeds the number of motion patterns (e.g., more than 100); and the datasets may have imbalanced distributions. To address these challenges, we introduce a new deep learning model called DeepeST (Deep Learning for Sub-Trajectory classification), which employs embedding vectors inspired by natural language processing techniques to manage large data volumes and tackle sparsity from subtrajectories. To our knowledge, DeepeST is the first model designed to address the challenges posed by imbalanced datasets in the TUL problem. We evaluated DeepeST’s performance through three case studies, comparing it against machine learning algorithms such as Random Forest and XGBoost and state-of-the-art deep learning approaches including MARC, BITULER, and TULVAE, using GPS and LBSN datasets for trajectory-user linking and criminal activities. DeepeST demonstrated significantly higher balanced accuracy, precision, recall, and F1-score values in all experiments.