Rose Yu "Learning from Large-Scale Spatiotemporal Data"

This lecture presents advanced deep learning methodologies for analyzing large-scale spatiotemporal data, which exhibits complex nonlinear correlations across both space and time that violate traditional machine learning assumptions of independence. The presenter introduces diffusion convolutional recurrent neural networks (DCRNNs) that integrate graph-based spatial modeling with temporal dynamics, enabling accurate forecasting of phenomena like urban traffic patterns. Key innovations include: (1) diffusion convolution operations that model how information propagates through non-Euclidean graph structures, (2) gated recurrent unit architectures that address vanishing gradient problems in temporal modeling, and (3) sequence-to-sequence frameworks with scheduled sampling for improved forecasting accuracy. The approach has demonstrated practical success in real-world applications including traffic prediction in Los Angeles and Bay Area networks, achieving forecasting accuracy up to one hour ahead—significantly outperforming traditional time series models and enabling deployment in production systems for intelligent transportation management.