MSI-Multi-Step Interaction Networks forSpatial-Temporal Forecasting
Year: 2023
Type: other
Abstract: Spatial-temporal (ST) forecasting is a critical but inherently challenging task in real-world applications such as traffic prediction and urban monitoring1 . Existing models often struggle to capture complex, longrange dependencies that vary across different locations and time steps2 . This paper introduces MSDR: Multi-Step Dependency Relation Networks, a novel framework designed to address these challenges. MSDR explicitly utilizes hidden states from multiple previous time steps, employing a "dependency relation operation" within a GMSDR (Graph Multi-Step Dependency Relation) block to learn high-quality representations3 . This block is integrated into an Encoder-Decoder architecture 4and uses graph convolutions and attention mechanisms to model both spatial and temporal dependencies5 . We evaluated our model on four public benchmark datasets (NYC Citi Bike, NYC Taxi, PEMS03, and PEMS08). The experimental results demonstrate that GMSDR achieves state-of-the-art performance, consistently outperforming established baselines like DCRNN, STGCN, and GraphWaveNet on all key metrics (RMSE, MAE, and MAPE%)
Authors: Tirumalasetty, Prasanth, Kilari, Navys sree, Nidamanuri, Shiva, Lu, Jin
Language: English
Subfield: Transportation
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7108339202