Towards Adaptive, Scalable and Interpretable Spatial-Temporal Forecasting
- Publication Type:
- Thesis
- Issue Date:
- 2023
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
In recent years, the application of deep learning models in the realm of spatial-temporal forecasting has significantly outpaced traditional statistical learning methods such as ARMA, VAR, and STL. However, a notable drawback to these models is their 'black box' nature, leaving their internal workings largely uninterpretable. Furthermore, recent studies have exposed the limitations of deep learning models, particularly those within the Transformer family, which often fall short when compared to carefully crafted linear models in various situations. These revelations invite reconsideration of the indispensability of deep learning models in spatial-temporal forecasting.
This thesis seeks to explore the complexities of modeling spatial-temporal data, striving to devise an interpretable, adaptive, and scalable spatial-temporal forecasting model that can compete with current state-of-the-art methodologies. By accomplishing this, it is anticipated that a deeper understanding of the underlying processes will be achieved, potentially paving the way for advancements in the field of spatial-temporal forecasting. This research encapsulates three key contributions: 1. The development of an attention-based method to probe correlations within diverse spatial-temporal contexts, addressing both adaptability and interpretability challenges; 2. The invention of a normalization-based method to take advantage of the characteristics of spatial and temporal data distribution, grappling with adaptability, interpretability, and scalability issues; 3. The introduction of a novel framework designed to emulate the dynamic behavior of underlying components in time series data, tackling adaptability, interpretability, and scalability concerns.
Please use this identifier to cite or link to this item: