Please use this identifier to cite or link to this item:
|Title:||Differentiable generative models for trajectory data analytics||Authors:||Li, Xiucheng||Keywords:||Engineering::Computer science and engineering::Mathematics of computing::Probability and statistics||Issue Date:||2019||Publisher:||Nanyang Technological University||Source:||Li, X. (2019). Differentiable generative models for trajectory data analytics. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||With the proliferation of GPS-enabled devices, trajectory data is being generated at an unprecedented speed. The trajectories are typically represented as sequences of discrete sample points, which carry rich spatiotemporal information. Mining patterns and distilling knowledge from such a large amount of trajectory data could potentially help address many real-world problems and improve our daily life experience. For instance, accurately and eﬃciently quantifying the similarity between two trajectories is a foundation for many trajectory based applications, such as tracking migration patterns of animals, mining hot routes in cities, trajectory clustering and moving group discovery. In this dissertation, we seek to eﬀectively and eﬃciently distill knowledge from trajectory data with diﬀerentiable generative models. We develop three ﬂexible generative models with eﬃciency in mind. The resulting models not only are capable of revealing the useful patterns underlying the data, but also admit end-to-end training. Moreover, our methods scale to real-world large-scale trajectory datasets easily. Speciﬁcally, we explore three important research problems arising in big trajectory data analytics: 1) learning representation for trajectory similarity computation; 2) learning the travel time distribution for any route on the road network; 3) spatial transition learning on the road network. In the study of trajectory representation learning, we propose the ﬁrst deep learning approach – t2vec – to learning representations of trajectories that is robust to low data quality, thus supporting accurate and eﬃcient trajectory similarity computation and search. Experiments show that our method is capable of higher accuracy and is at least one order of magnitude faster than the state-of-the-art methods for k-nearest trajectory search. In the study of travel time distribution learning, we develop a novel deep generative model – DeepGTT – to learn the travel time distribution for any route on the route network by conditioning on the real-time traﬃc. DeepGTT interprets the generation of travel time using a three-layers hierarchical probabilistic model, and describes the generation process in a reasonable manner rather than simply learning by brute force, and thus it not only produces more accurate results but also is quite data-eﬃcient. A variational loss is further derived and the entire model is fully diﬀerentiable, which makes the model easily scale to large data sets. In the study of spatial transition learning on the road network, we present a novel deep probabilistic model – DeepST – which uniﬁes three explanatory factors, the past traveled route, the impact of destination and real-time traﬃc for the route decision. DeepST explains the generation of next road link by conditioning on the representations of the three explanatory factors. To enable eﬀectively sharing the statistical strength, we propose to learn representations of k-destination proxies with an adjoint generative model. To incorporate the impact of real-time traﬃc, we introduce a high-dimensional latent variable as its representation whose posterior distribution can then be inferred from observations. An eﬃcient inference method is developed within the Variational Auto-Encoders framework to scale DeepST to large-scale data sets.||URI:||https://hdl.handle.net/10356/137159||DOI:||10.32657/10356/137159||Rights:||This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Updated on Feb 7, 2023
Updated on Feb 7, 2023
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.