Spatial-Temporal Representation Learning for Traffic Forecasting
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Traffic forecasting is the cornerstone of the Intelligent Transportation System (ITS). Accurate traffic prediction can effectively promote traffic management and urban planning. This task is challenging due to the dynamic and complex spatial-temporal correlations in the time-evolving traffic network. Spatially, the traffic conditions of nearby locations have a dynamic influence on each other. Temporally, the traffic conditions exhibit elusive patterns due to various external factors such as weather, rush hour, weekends, holidays, etc. With the development of deep learning, spatial-temporal representation learning has become the mainstream approach to traffic forecasting tasks. This thesis investigates techniques for learning effective spatial-temporal representation for traffic forecasting systems. The exploration is motivated by the following challenges that could hinder the further development of spatial-temporal representation learning for enhancing the final forecasting performance.
1. The diverse complexity of spatial-temporal representation learning. The complexity of diverse forecasting tasks is non-uniformly distributed across various spatial locations (e.g. suburb vs. downtown) and temporal steps (e.g. rush hour vs. off-peak).
2. The data scarcity and imbalance in traffic forecasting datasets. For example, the temporal observations of traffic accidents exhibit ultra-rareness due to the inherent properties of accident occurrences. This leads to the severe scarcity of risk samples for comprehensive representation learning.
3. The under-exploration of frequency domain spatial-temporal representation learning. The spectral features could provide potential compensation for the time domain representation learning.
4. Distribution shift in the non-stationary traffic data. Despite the recent success of deep neural networks in spatial-temporal forecasting, existing methods suffer from distribution shifts between the training and test data, failing to address the non-stationary and abrupt changes at test time.
To solve the above challenges, this thesis proposes four models for various traffic forecasting applications. All of these methods aim to explore challenging spatial-temporal representation learning for enhancing traffic forecasting performance. Specifically, we first propose a Bidirectional Spatial-Temporal Adaptive Transformer (Bi-STAT) for accurate traffic flow forecasting, which devises the recurrent mechanism with a novel Dynamic Halting Module (DHM) to dynamically learn the spatial-temporal representation in the traffic streams according to their complexities. Secondly, we propose a contrastive learning approach with the multi-kernel networks, to learn the traffic accident representation under temporal scarcity and imbalanced spatial distribution. Then, we design a novel embedded 2D spectral learning framework to explore the traffic features in the frequency domain. Lastly, we propose a novel test-time training framework for spatial-temporal representation learning to alleviate the distribution shift in traffic data.
Please use this identifier to cite or link to this item: