Leveraging Siamese Framework for Enhanced Contextual Feature Learning in Traffic Forecasting

As smart city systems become increasingly complex and urban road networks become more interconnected, traditional predictive models are gradually losing their ability to handle the intricate information among contextual data, such as traffic flow and weather data, due to their inability to learn the holistic features of the entire dataset. These data not only require consideration of contextual relationships but also an understanding of the overall characteristics of the data, which traditional models typically fail to achieve. To address these challenges, this paper proposes a novel deep learning architecture: SimFormer. This architecture leverages the advantages of both the Siamese network and Transformer architectures, mapping complex time series into high-dimensional space to better capture their contextual features. Concurrently, the Transformer encoder's focus on capturing global temporal dynamics ensures that global features are preserved and utilized in subsequent prediction tasks. We evaluate this innovative architecture on multiple real-world datasets using two different performance metrics, and it consistently outperforms traditional models.