Deep Learning Based Feature Engineering for Discovering Spatio-Temporal Dependency in Traffic Flow Forecasting

Abstract

Intelligent transportation systems (ITS) have garnered considerable attention for providing efficient traffic management solutions. Traffic flow forecasting is a crucial component of it which serves as the foundation for various state-of-the-art deep learning approaches. Initially, researchers recognized that significant temporal changes from traffic flow data for modelling. However, as researchers delved deeper into the underlying correlations within traffic flow data, they discovered that spatial information from the road network also plays a crucial role in accurate forecasting. Consequently, deep learning methods that incorporate Spatio-temporal representation have been employed to address traffic flow forecasting. Although recent solutions to this problem are impressive, it is essential to discuss the reasoning behind the architecture of the model. The expression of each feature relies on selecting appropriate models for feature extraction and designing architectures that minimize information loss during modeling. In this thesis, the work focuses on graph-based Spatio-temporal feature engineering. The experiments are divided into two parts: 1). explores the efficient architecture for expressing spatial-temporal information by considering both different sequential modelling approaches. 2). Based on the result obtained, the second experiment focuses on multi- scale modelling to capture informative Spatio-temporal feature. We propose a model that incorporates sequential modeling and captures multi-scale Spatiotemporal semantics by employing residual connections in different hierarchy. We validate our model using three datasets, each containing varying information for extraction. Taking into account the dataset characteristics and the model structure, our model outperforms the baselines and state-of-the-art models. The experimental results indicate that the performance of sequential modeling and multi-scale semantics, combined with thoughtful model design, significantly contribute to the overall forecasting performance. Furthermore, our work serves as inspiration for expressive data mining methods that rely on appropriate feature extraction models and architecture design, taking into consideration the information content within the dataset.